NeurIPS2025 NeurIPS2025 accepted papers NeurIPS2025 paper list AI paper notes top conference papers Image Generation Safety Model Compression Reinforcement Learning Optimization & Theory Multimodal VLM LLM Reasoning Interpretability

🧠 NeurIPS2025 Accepted Papers¶

2530 NeurIPS2025 paper notes covering Image Generation (222), Safety (156), Model Compression (146), Reinforcement Learning (143), Optimization & Theory (127), Multimodal VLM (108), LLM Reasoning (82), Interpretability (81) and other 53 areas. Each note has TL;DR, motivation, method, experiments, highlights, and limitations — 5-minute reads of core ideas.

💡 LLM Reasoning (82)¶

The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity: Using controlled puzzle environments, this paper systematically reveals a three-regime behavioral pattern in Large Reasoning Models (LRMs): performance falls below standard LLMs at low complexity (overthinking), substantially surpasses them at moderate complexity, and collapses completely (0%) at high complexity. Counterintuitively, models reduce thinking token usage at the point of collapse, demonstrating that current LRMs have not developed genuinely generalizable reasoning capabilities.
Note 4: WebThinker — Empowering Reasoning Models with Deep Research Capabilities: WebThinker equips large reasoning models (LRMs) with autonomous web search and navigation capabilities. Through a Think-Search-Draft strategy, it seamlessly interleaves reasoning, information gathering, and report generation. After reinforcement learning optimization, it surpasses o1 and Gemini on complex reasoning and scientific report generation tasks.
SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution: This work is the first to apply reinforcement learning (RL) to real-world software engineering tasks (GitHub PR/Issue resolution), training Llama-3.3-70B exclusively with a rule-based sequence-similarity reward. It achieves a 41.0% resolve rate on SWE-bench Verified (SOTA among medium-scale models). Notably, although RL training is conducted solely on issue-solving data, it elicits emergent generalization in out-of-domain tasks including code reasoning, mathematics, and general language understanding.
Scalable Best-of-N Selection for Large Language Models via Self-Certainty: This paper proposes Self-Certainty, a metric that quantifies model confidence via the token probability distribution of LLM outputs, enabling scalable Best-of-N selection without any auxiliary reward model. The approach achieves performance comparable to or exceeding reward-model-based methods.
The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning: This paper decomposes reinforcement learning with verifiable rewards (RLVR) into positive sample reinforcement (PSR, which increases the probability of correct responses) and negative sample reinforcement (NSR, which penalizes incorrect responses). It finds that NSR alone consistently improves reasoning performance across the full Pass@k spectrum and typically matches or surpasses PPO/GRPO. Based on this finding, the paper proposes Weighted-REINFORCE (reducing the PSR weight to 0.1), achieving state-of-the-art results across MATH, AIME 2025, and AMC23.
Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning: This paper demonstrates that excessively extending CoT length degrades LLM reasoning performance, and proposes Thinking-Optimal Scaling (TOPS), a strategy that trains models to select the shortest correct response for each problem via self-improvement, outperforming existing distillation methods in both accuracy and efficiency.
VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning: This paper proposes VideoRFT, which extends the reinforced fine-tuning (RFT) paradigm to video reasoning via a cognition-inspired multi-expert CoT data construction pipeline and a novel semantic consistency reward. Two datasets are constructed: VideoRFT-CoT-102K (for SFT) and VideoRFT-RL-310K (for RL), achieving state-of-the-art performance on 6 video reasoning benchmarks.
SQL-R1: Training Natural Language to SQL Reasoning Model By Reinforcement Learning: This work presents the first systematic application of GRPO-based reinforcement learning to NL2SQL tasks. Through a four-level progressive reward function and a training strategy combining 200K cold-start data with 5K complex-sample RL fine-tuning, the 7B model achieves 88.7% on Spider and 66.6% on BIRD, surpassing GPT-4-based methods at comparable scale.
Reasoning Models Better Express Their Confidence: This paper systematically demonstrates that reasoning models (with extended CoT) exhibit significantly better confidence calibration than non-reasoning models, and identifies "slow-thinking" behaviors—exploring alternatives, backtracking, and verification—as the fundamental source of this calibration improvement.
Segment Policy Optimization: Effective Segment-Level Credit Assignment in RL for Large Language Models: This paper proposes the SPO framework, which adopts segment-level (rather than token-level or trajectory-level) advantage estimation. Through a novel Monte Carlo method and tree-based sampling, SPO outperforms PPO and GRPO by 6–12 and 7–11 percentage points in short-CoT and long-CoT settings, respectively.

Browse all 82 LLM Reasoning papers →

🦾 LLM Agent (56)¶

A-MEM: Agentic Memory for LLM Agents: This paper proposes A-Mem, a Zettelkasten-inspired agentic memory system for LLM agents. Each memory entry automatically generates a structured note (keywords/tags/contextual description), dynamically establishes inter-memory links, and triggers evolutionary updates to existing memories upon the insertion of new ones. A-Mem substantially outperforms baselines such as MemGPT on the LoCoMo long-conversation QA benchmark.
Group-in-Group Policy Optimization for LLM Agent Training: GiGPO introduces step-level grouping nested within the episode-level grouping of GRPO by leveraging recurring environment states across trajectories as anchor states, enabling fine-grained credit assignment without additional rollouts or a critic model. It outperforms GRPO by >12% on ALFWorld and >9% on WebShop.
Multi-Agent Collaboration via Evolving Orchestration: This paper proposes a "Puppeteer" multi-agent collaboration paradigm in which a centralized orchestrator learns via RL to dynamically select which agent to activate at each reasoning step. The approach simultaneously improves performance and efficiency on both closed-domain and open-domain tasks, and reveals that evolved topologies tend toward more compact cyclic structures.
AgentAuditor: Human-Level Safety and Security Evaluation for LLM Agents: This paper proposes AgentAuditor — a training-free, memory-augmented reasoning framework that enables LLMs to adaptively extract structured semantic features (scenario, risk, behavior) to construct an experiential memory bank, then employs multi-stage context-aware retrieval-augmented generation to guide LLM evaluators in assessing agent behavior for safety and security threats. The work also introduces ASSEBench, the first benchmark jointly covering safety and security evaluation (2,293 records, 15 risk types, 29 scenarios), achieving human expert-level evaluation accuracy across multiple benchmarks.
Deep Video Discovery: Agentic Search with Tool Use for Long-form Video Understanding: This paper proposes DVD (Deep Video Discovery), an agent that frames long-form video understanding as a multi-step information search problem. It first constructs a multi-granular structured database from a long video (global summary + clip-level caption embeddings + frame-level pixels), then provides three search tools (Global Browse / Clip Search / Frame Inspect). A reasoning LLM autonomously orchestrates the search trajectory via an observe-reason-act loop. DVD achieves 74.2% on LVBench (surpassing the previous SOTA MR.Video by 13.4 pp), and 76.0% with subtitles.
Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models?: This work establishes, both theoretically and empirically, that the performance gains attributed to Multi-Agent Debate (MAD) stem primarily from majority voting (ensembling) rather than the debate process itself. The debate dynamics are shown to constitute a martingale—meaning debate does not systematically improve correctness in expectation—and this theoretical insight motivates a principled improvement to MAD by biasing updates toward correct signals.
ContextAgent: Context-Aware Proactive LLM Agents with Open-World Sensory Perceptions: This paper proposes ContextAgent, the first LLM agent framework that leverages multimodal sensory perception from wearable devices (video + audio + notifications) to understand user intent and proactively deliver tool-augmented services. It also introduces ContextAgentBench, a benchmark of 1,000 samples, achieving improvements of 8.5% in proactive prediction accuracy and 6.0% in tool invocation accuracy.
Distilling LLM Agent into Small Models with Retrieval and Code Tools: This paper proposes an Agent Distillation framework that distills the complete reason-act-observe interactive behaviors of LLM agents (rather than static CoT) into small models ranging from 0.5B to 7B parameters. Combined with a first-thought prefix to improve teacher trajectory quality and self-consistent action generation to enhance inference robustness, the framework enables small models to achieve performance comparable to CoT-distilled models 2–4× their size.
MAT-Agent: Adaptive Multi-Agent Training Optimization: This paper proposes MAT-Agent, a multi-agent framework consisting of four autonomous agents responsible for data augmentation, optimizer selection, learning rate scheduling, and loss function selection, respectively. The framework dynamically adjusts training configurations during the training process, employing DQN to learn policies as a replacement for conventional static hyperparameter configurations, and achieves state-of-the-art performance on multi-label image classification tasks.
MedAgentBoard: Benchmarking Multi-Agent Collaboration with Conventional Methods for Diverse Medical Tasks: This paper proposes MedAgentBoard, a comprehensive benchmark that systematically evaluates multi-agent collaboration, single-LLM, and conventional methods across diverse medical tasks, revealing that multi-agent collaboration does not consistently outperform strong single models or specialized conventional approaches.

Browse all 56 LLM Agent papers →

⚖️ Alignment & RLHF (36)¶

Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond): This work introduces the Infinity-Chat dataset (26K open-ended real-world user queries with 31,250 human annotations) to expose the "Artificial Hivemind" phenomenon in language models — severe intra-model repetition and inter-model homogeneity in open-ended generation — and demonstrates that Reward Models and LM Judges fail to calibrate on samples with high inter-annotator preference divergence.
What Makes a Reward Model a Good Teacher? An Optimization Perspective: From an optimization-theoretic perspective, this paper proves that reward model accuracy alone is insufficient to measure its quality as an RLHF "teacher." Even a perfectly accurate reward model can lead to a flat RLHF objective landscape and extremely slow policy gradient optimization if the induced reward variance is too low. Moreover, different language models require different reward models.
DeepVideo-R1: Video Reinforcement Fine-Tuning via Difficulty-aware Regressive GRPO: This paper proposes DeepVideo-R1, which reformulates GRPO as Reg-GRPO that directly regresses advantage values (eliminating clipping/min safeguards), and mitigates the vanishing advantage problem via difficulty-aware data augmentation, achieving up to 10.1 percentage points improvement over standard GRPO on video reasoning tasks.
Improving Data Efficiency for LLM Reinforcement Fine-tuning Through Difficulty-targeted Online Data Selection and Rollout Replay: Two complementary techniques are proposed to improve the data efficiency of LLM reinforcement fine-tuning (GRPO): (1) DOTS — an attention-based mechanism for predicting adaptive difficulty that prioritizes moderate-difficulty questions to maximize gradient signal; and (2) Rollout Replay — reusing recent rollouts to reduce per-step computational overhead. Together, these techniques reduce training time by an average of 40.7% across 6 model–dataset combinations.
Reinforcement Learning Finetunes Small Subnetworks in Large Language Models: RL fine-tuning of LLMs updates only 5%–30% of parameters in practice (sparse subnetworks), and these subnetworks exhibit high consistency across different random seeds, datasets, and algorithms. Fine-tuning only the identified subnetwork can reproduce both the performance and the parameter values of full fine-tuning.
GVPO: Group Variance Policy Optimization for Large Language Model Post-Training: GVPO is a more stable LLM post-training method than GRPO, derived by embedding the analytical solution of KL-constrained reward maximization into gradient weights (zero-sum weights eliminate the partition function). It achieves 20.72% on AIME (vs. GRPO's 14.79%) and is proven to possess a unique global optimum.
Trajectory Balance with Asynchrony: Decoupling Exploration and Learning for Fast, Scalable LLM Post-Training: This paper proposes TBA (Trajectory Balance with Asynchrony), which combines the GFlowNet Trajectory Balance (TB) objective with an asynchronous distributed RL architecture to decouple exploration and learning in LLM post-training, achieving 4–50× speedups without performance degradation across mathematical reasoning, preference fine-tuning, and automated red-teaming tasks.
Mechanism Design for LLM Fine-tuning with Multiple Reward Models: This paper formulates multi-party preference aggregation in RLHF fine-tuning as a mechanism design problem. It proves that under social-welfare-maximizing training rules, participants have incentives to misreport their preferences, and achieves dominant-strategy incentive compatibility (DSIC) via an extended VCG payment mechanism that ensures truthful reporting.
From Judgment to Interference: Early Stopping LLM Harmful Outputs via Streaming Content Monitoring: This paper proposes the Streaming Content Monitor (SCM)—the first harmful content monitor natively designed for partial detection. Built upon the FineHarm dataset (29K samples with token-level annotations) and hierarchical consistency-aware learning, SCM achieves a macro F1 of 0.95+ after observing on average only 18% of response tokens, enabling real-time early stopping of harmful LLM outputs.
Adjacent Words, Divergent Intents: Jailbreaking Large Language Models via Task Concurrency: This paper proposes JAIL-CON, a jailbreak attack framework based on task concurrency. By interleaving harmful and benign tasks at the word level, it exploits LLMs' ability to handle concurrent tasks to bypass safety mechanisms, while the resulting concurrent outputs exhibit stronger evasiveness against guardrails.

Browse all 36 Alignment & RLHF papers →

🔒 Safety (156)¶

Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning: This paper identifies that reference model bias in NPO (Negative Preference Optimization) leads to uneven optimization power allocation across forget data and early-stage gradient weight smoothing failure. The proposed SimNPO eliminates reference model dependency and adopts length-normalized rewards, improving FQ from 0.79 to 0.99 on TOFU and consistently outperforming NPO across all benchmarks.
AgentDAM: Privacy Leakage Evaluation for Autonomous Web Agents: This paper proposes AgentDAM, the first benchmark for end-to-end evaluation of data minimization compliance by AI agents in real web environments. It comprises 246 tasks spanning Reddit, GitLab, and Shopping platforms, and finds that leading models such as GPT-4o exhibit privacy leakage rates of 36–46% without mitigation, while a CoT-based privacy prompt reduces leakage rates to 6–8%.
DRIFT: Dynamic Rule-Based Defense with Injection Isolation for Securing LLM Agents: DRIFT is a system-level agent security framework featuring three layers of defense: Secure Planner (pre-planned function trajectories and parameter checklists), Dynamic Validator (dynamic policy updates based on Read/Write/Execute permissions), and Injection Isolator (detection and masking of injected instructions from the memory stream). On AgentDojo, DRIFT reduces ASR from 30.7% to 1.3% while achieving 20.1% higher utility than CaMeL.
Contextual Integrity in LLMs via Reasoning and Reinforcement Learning: This paper proposes CI-RL, a framework that combines Chain-of-Thought reasoning prompts with GRPO reinforcement learning to train LLMs to understand contextual integrity (CI) using only ~700 synthetic samples. On the PrivacyLens benchmark, it reduces privacy leakage rates by up to 40%, and smaller models trained with CI-RL can surpass larger baseline models.
Attractive Metadata Attack: Inducing LLM Agents to Invoke Malicious Tools: AMA (Attractive Metadata Attack) demonstrates that by carefully crafting malicious tool metadata (name, description, parameter schema) alone — without prompt injection or internal model access — an attacker can induce LLM agents to invoke malicious tools and leak private data at a success rate of 81–95%, while barely affecting original task completion (98%+), with existing defenses (auditors, prompt rewriting) proving largely ineffective.
Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text: This paper proposes Adversarial Paraphrasing — a training-free universal attack framework that selects the most "human-like" token at each decoding step by leveraging feedback signals from AI text detectors during token-by-token paraphrasing. The approach achieves an average T@1%F reduction of 87.88% across 8 detectors and exhibits strong cross-detector transferability.
ReliabilityRAG: Effective and Provably Robust Defense for RAG-based Web-Search: ReliabilityRAG proposes a RAG framework that leverages document reliability signals (e.g., search ranking) for adversarial defense. It identifies a consistent document subset by finding the Maximum Independent Set (MIS) on a contradiction graph while prioritizing high-reliability documents, providing provable robustness guarantees alongside high accuracy on benign scenarios and long-form generation tasks.
ForensicHub: A Unified Benchmark & Codebase for All-Domain Fake Image Detection and Localization: ForensicHub introduces the first unified benchmark platform spanning all domains (Deepfake/IMDL/AIGC/Document Tampering) for fake image detection and localization, encompassing 4 tasks, 23 datasets, 42 models, 6 backbone networks, and 11 GPU-accelerated evaluation metrics. Through a modular architecture and adapter design, it bridges domain silos and conducts 16 cross-domain evaluations to derive 8 key insights.
LLM Strategic Reasoning: Agentic Study through Behavioral Game Theory: This paper proposes an LLM strategic reasoning evaluation framework grounded in behavioral game theory. It employs Truncated Quantal Response Equilibrium (TQRE) to quantify reasoning depth τ, evaluates 22 state-of-the-art models across 13 matrix games, and reveals differences in reasoning styles as well as biases induced by demographic personas.
Rewind-to-Delete: Certified Machine Unlearning for Nonconvex Functions: This paper proposes R2D (Rewind-to-Delete), the first first-order, black-box certified machine unlearning algorithm for general nonconvex loss functions. It achieves data deletion by rewinding to an earlier checkpoint in the training trajectory and then performing gradient descent on the retained data, while providing $(ε, δ)$-certified unlearning guarantees and theoretical trade-offs among privacy, utility, and efficiency.

Browse all 156 Safety papers →

👻 Hallucination Detection (17)¶

Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling: This paper proposes REVERSE, the first framework to unify generation adjustment and post-hoc verification within a single VLM. Through hallucination-aware training on 1.3M semi-synthetic samples combined with inference-time retrospective resampling, REVERSE enables a VLM to automatically detect and correct hallucinations during generation, achieving a 12% reduction on CHAIR-MSCOCO and a 34% improvement on HaloQuest.
Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models: This paper systematically audits the generation and propagation mechanisms of hallucinations in reasoning large language models (RLLMs), finding that reflection in long CoT amplifies hallucinations through metacognitive bias rather than correcting them. Even targeted interventions at the hallucination source fail to alter final outputs (chain disloyalty), exposing critical shortcomings of existing hallucination detection methods in multi-step reasoning scenarios.
Robust Hallucination Detection in LLMs via Adaptive Token Selection: HaMI frames hallucination detection as a Multiple Instance Learning (MIL) problem, treating each generated sequence as a bag of token instances. By jointly optimizing token selection and hallucination detection, it adaptively identifies the most informative tokens, achieving substantial AUROC improvements over all existing methods across four QA benchmarks (up to 11.9%).
Mitigating Hallucination Through Theory-Consistent Symmetric Multimodal Preference Optimization: This paper proposes SymMPO (Symmetric Multimodal Preference Optimization), which addresses two key limitations of existing vision-augmented DPO methods—namely, theoretically unsound objective functions and indirect preference supervision—through symmetric paired preference learning over contrastive images and preference margin consistency regularization. Consistent performance gains are achieved across five hallucination benchmarks.
Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT: This paper organizes all hidden-layer activations of an LLM into an "activation tensor" (layers × tokens × hidden dimension), treats it analogously to an image, and processes it with a ViT-based architecture (ACT-ViT) that supports joint training across multiple LLMs. The method consistently outperforms conventional probing approaches across 15 LLM–dataset combinations and demonstrates strong zero-shot/few-shot transfer to unseen datasets and unseen LLMs.
When Semantics Mislead Vision: Mitigating Large Multimodal Models Hallucinations: This paper identifies a "semantic hallucination" problem in Large Multimodal Models (LMMs) for scene text recognition—where non-semantic text is misread as semantically plausible words. Analysis reveals that Transformer layers whose attention is more focused on text regions are less prone to hallucination. Based on this finding, the authors propose a training-free framework, ZoomText + Grounded Layer Correction, achieving approximately 4–5% improvement on TextHalu-Bench and approximately 4% on ST-VQA.
Seeing is Believing? Mitigating OCR Hallucinations in Multimodal Large Language Models: This paper addresses OCR hallucinations in MLLMs under degraded document conditions. It introduces KIE-HVQA, the first benchmark for evaluating hallucinations in degraded document scenarios, and proposes a multi-objective reward reinforcement learning framework based on GRPO. The resulting 7B-parameter model achieves approximately 28% higher hallucination-suppression accuracy than GPT-4o.
GLSim: Detecting Object Hallucinations in LVLMs via Global-Local Similarity: GLSim is a training-free object hallucination detection method for LVLMs that combines a global scene similarity score (cosine similarity between the object token and the last instruction token) and a local visual grounding similarity score (cosine similarity between the object token and the Top-K image patch embeddings localized via Visual Logit Lens). It achieves 83.7% AUROC on MSCOCO, surpassing SVAR by 9% and Internal Confidence by 10.8%.
SECA: Semantically Equivalent and Coherent Attacks for Eliciting LLM Hallucinations: This paper proposes SECA (Semantically Equivalent and Coherent Attacks), a realistic prompt perturbation framework that elicits LLM hallucinations while preserving semantic equivalence and coherence, achieving higher attack success rates on multiple-choice QA tasks with near-zero semantic errors.
Generalization or Hallucination? Understanding Out-of-Context Reasoning in Transformers: This paper argues that LLM generalization and hallucination share a common mechanism — out-of-context reasoning (OCR) — and provides theoretical guarantees on a single-layer attention model: the factorized parameterization $(W_O, W_V)$ can perform OCR due to the nuclear norm implicit bias of gradient descent, whereas the merged parameterization $W_{OV}$ cannot due to its Frobenius norm bias. Moreover, OCR is sample-efficient (requiring only $m_{\text{train}}>0$).

Browse all 17 Hallucination Detection papers →

📊 LLM Evaluation (38)¶

ConfTuner: Training Large Language Models to Express Their Confidence Verbally: ConfTuner proposes a tokenized Brier score loss function (theoretically proven to be a proper scoring rule) that requires only 2000 samples and 4 minutes of LoRA fine-tuning to enable LLMs to output calibrated verbalized confidence (e.g., "I am 80% confident"), reducing ECE by up to 60.9% and supporting downstream applications like self-correction and model cascading.
ConTextTab: A Semantics-Aware Tabular In-Context Learner: ConTextTab integrates semantic embeddings (text encodings of column names and categorical values) into a table-native ICL architecture, and pretrains on large-scale real-world tabular data (T4, ~2.18M tables). It achieves a new state of the art on the semantics-rich CARTE benchmark while remaining competitive with existing methods on non-semantic benchmarks.
PaTH Attention: Position Encoding via Accumulating Householder Transformations: This paper proposes PaTH (Position encoding via accumulating Householder Transformations), a data-dependent multiplicative position encoding scheme that replaces RoPE's static rotation matrices with accumulated Householder transformations, achieving superior theoretical expressiveness and empirical language modeling performance over RoPE.
OptiTree: Hierarchical Thoughts Generation with Tree Search for LLM Optimization Modeling: This paper proposes OptiTree, which organizes hierarchical classification and modeling thoughts for operations research (OR) problems by constructing a modeling tree, and employs tree search to adaptively decompose complex problems into sequences of simpler subproblems, achieving significant accuracy gains in optimization modeling tasks for LLMs (exceeding 10% on multiple challenging benchmarks).
Hyperbolic Fine-Tuning for Large Language Models: This work identifies that LLM token embeddings follow power-law distributions and exhibit tree-like hyperbolic structure, and proposes HypLoRA — performing low-rank adaptation directly on the Lorentz hyperbolic manifold (bypassing the cancellation effect of tangent space mappings) — achieving significant gains over standard LoRA on arithmetic and commonsense reasoning tasks (e.g., M.AVG +7.5% on Qwen2.5-7B).
Time Travel is Cheating: Going Live with DeepFund for Real-Time Fund Investment Benchmarking: This paper introduces DeepFund — the first live fund investment benchmark for LLMs — which employs a multi-agent architecture (Financial Planner + Analyst Team + Portfolio Manager) connected to real-time market data, eliminating the information leakage caused by LLM "time travel" in traditional backtesting. Over 24 trading days of live testing across 9 flagship LLMs, only Grok 3 achieves positive returns, revealing fundamental limitations of current LLMs in active fund management.
HybridNorm: Towards Stable and Efficient Transformer Training via Hybrid Normalization: This paper proposes HybridNorm, a hybrid normalization strategy that applies QKV normalization within the attention module to decouple gradients and Post-Norm within the FFN to enhance regularization. Across scales from 550M to 7B parameters, HybridNorm simultaneously achieves the training stability of Pre-Norm and the generalization performance of Post-Norm, yielding an average downstream task improvement of 2.45% at the 7B scale.
Your Pre-trained LLM is Secretly an Unsupervised Confidence Calibrator: This paper identifies that post-training (SFT/RLHF/DPO) degrades the confidence calibration of pre-trained language models, and proposes DACA, a method that exploits the well-calibrated nature of pre-trained models by aligning confidence distributions exclusively on prediction-consistent samples, achieving label-free calibration of post-trained models with up to 15.08% ECE improvement.
Toward Engineering AGI: Benchmarking the Engineering Design Capabilities of LLMs: This paper introduces EngDesign—the first LLM engineering design benchmark spanning 9 engineering domains (operating systems, computer architecture, control systems, mechanical engineering, structural engineering, digital hardware, analog circuits, robotics, and signal processing)—replacing conventional QA matching with a simulation-driven evaluation pipeline. The benchmark reveals that even the most capable reasoning model, o3, achieves only a 34% pass rate.
ComPO: Preference Alignment via Comparison Oracles: To address likelihood displacement and verbosity caused by noisy preference pairs (where preferred and dispreferred responses are highly similar) in DPO, this paper proposes ComPO, a zeroth-order preference alignment method based on comparison oracles. The approach partitions data into clean and noisy subsets, applying DPO to the clean subset and ComPO to extract alignment signals from the noisy subset, achieving consistent improvements in LC win rate on benchmarks such as AlpacaEval 2.

Browse all 38 LLM Evaluation papers →

⚡ LLM Efficiency (34)¶

Tensor Product Attention Is All You Need: By decomposing Q/K/V into weighted sums of low-rank factors via contextual tensor products, TPA compresses the KV cache by 10–16×, while surpassing standard MHA/MQA/GQA/MLA on both validation loss and downstream task accuracy.
Advancing Expert Specialization for Better MoE: By jointly optimizing an orthogonality loss (reducing projection overlap among experts) and a variance loss (increasing routing score diversity), the proposed method reduces expert overlap by 45% and improves routing variance by 150% without modifying the MoE architecture, achieving an average gain of 23.79% across 11 benchmarks while fully preserving load balance.
Jet-Nemotron: Efficient Language Model with Post Neural Architecture Search: NVIDIA proposes the PostNAS pipeline — starting from a pretrained full-attention model, freezing MLP weights, and applying a four-step search (full-attention layer placement → linear attention block selection → novel JetBlock design → hardware-aware hyperparameter search) to yield the hybrid Jet-Nemotron architecture. The 2B model surpasses Qwen3-1.7B on MMLU-Pro while achieving 47× higher generation throughput.
The Emergence of Sparse Attention: Impact of Data Distribution and Benefits of Repetition: This paper investigates the emergence mechanism of sparse attention through theoretical analysis and controlled experiments, revealing that the emergence time follows a power-law relationship with respect to sequence length and dimensionality, $T_\epsilon \propto \sqrt{d} \cdot T$. It further demonstrates that both in-context and cross-sample data repetition strategies accelerate emergence, offering a unified sparse attention perspective for understanding capability emergence in LLMs.
Critical Batch Size Revisited: A Simple Empirical Approach to Large-Batch Language Model Training: This paper proposes a branched training method to directly measure the critical batch size (CBS) empirically, finding that CBS grows rapidly in early training before plateauing and is independent of model scale. Based on this insight, a batch size warmup strategy is designed that achieves equivalent or superior training loss with 43% fewer gradient steps.
Constant Bit-Size Transformers Are Turing Complete: This paper provides the first proof that a Transformer with constant bit-size precision and a fixed number of parameters — permitting only context window growth — is Turing complete. It establishes the exact complexity equivalence WINDOW[s(n)] = SPACE[s(n)], demonstrating that expanding the context window, rather than model size, suffices for universal computation.
MoESD: Revealing the Potential of Speculative Decoding to Accelerate Sparse MoE: This work challenges the prevailing belief that speculative decoding (SD) is ineffective for MoE models. Through theoretical analysis and experiments, it demonstrates that MoE models benefit more from SD than dense models at medium batch sizes. The paper introduces target efficiency as a system-level metric to quantify acceleration bottlenecks, constructs a reliable performance prediction model, and achieves up to 2.29× speedup on Qwen2-57B-A14B.
Tiled Flash Linear Attention: More Efficient Linear RNN and xLSTM Kernels: This paper proposes TFLA (Tiled Flash Linear Attention), which achieves efficient linear RNN/mLSTM kernels through two-level sequence parallelism and tiling optimization, delivering significant wall-clock speedups over FlashAttention 3 and Mamba 2 (>2× in training vs. Mamba 2) while maintaining equivalent model accuracy.
L-MTP: Leap Multi-Token Prediction Beyond Adjacent Context for Large Language Models: L-MTP introduces a leap mechanism into multi-token prediction (MTP) by predicting tokens at non-adjacent positions (e.g., positions 1, 3, 5, 7 instead of 1, 2, 3, 4). A "looking backward" decoding strategy reuses prior predictions to fill the gaps, achieving a 22% inference speedup on 3B–12B models while maintaining or improving task performance.
Efficient Training-Free Online Routing for High-Volume Multi-LLM Serving: This paper proposes PORT, the first training-free online LLM routing algorithm. PORT estimates query features via approximate nearest neighbor search (ANNS) and performs a one-shot optimization of dual variables as routing weights on a small set of initial queries. Under a limited token budget, PORT achieves near-offline-optimal routing performance with a $1-o(1)$ competitive ratio, delivering on average 3.55× performance improvement, 1.85× cost efficiency, and 4.25× throughput over baselines.

Browse all 34 LLM Efficiency papers →

📚 Pretraining (52)¶

Nemotron-CLIMB: CLustering-based Iterative Data Mixture Bootstrapping for Language Model Pre-training: NVIDIA proposes the CLIMB framework, which automatically discovers optimal pre-training data mixture ratios through embedding-based clustering and iterative bootstrapped search. On a 1B-scale model, CLIMB outperforms Llama-3.2-1B by 2.0%, and releases the 1.2T-token ClimbLab corpus and the 400B-token ClimbMix high-quality dataset.
The Curse of Depth in Large Language Models: This paper identifies the root cause of deep-layer degradation in Pre-LN Transformers—exponential growth of output variance causing deep layers to collapse into identity mappings—and proposes a parameter-free LayerNorm Scaling (LNS) strategy that multiplies the LayerNorm output by $1/\sqrt{\ell}$, compressing variance growth from exponential to polynomial. LNS consistently improves perplexity by 5–8% across scales from 130M to 7B parameters.
Power Lines: Scaling Laws for Weight Decay and Batch Size in LLM Pre-training: This paper proposes a set of power-law scaling relations for weight decay $\lambda$ and batch size $B$ in LLM pre-training. By introducing the concept of an AdamW timescale $\tau$, it unifies hyperparameter scaling relationships, enabling accurate prediction of optimal hyperparameters prior to large-scale training.
Scaling Embedding Layers in Language Models: This paper proposes Scone, a method that learns contextualized embeddings for high-frequency n-grams using a separate Transformer model, and offloads these embeddings to main memory/SSD at inference time. This enables a new scaling paradigm in which additional compute is consumed during training without increasing accelerator resource usage at inference. A 1B-parameter Scone model surpasses a 1.9B baseline.
Does Object Binding Naturally Emerge in Large Pretrained Vision Transformers?: By defining the IsSameObject predicate and designing quadratic probes, this work demonstrates that large-scale pretrained ViTs — particularly DINO and CLIP — naturally develop object binding capabilities. This signal is encoded in a low-dimensional subspace and actively guides the attention mechanism, challenging the cognitive science community's view that ViTs lack binding ability.
Broken Tokens: Your Language Model Can Secretly Handle Non-Canonical Tokenization: This paper reveals that LLMs can secretly handle non-canonical tokenizations (e.g., splitting "Hello" into "He"+"llo" instead of the canonical whole-word token)—even when the input token sequence differs from training, models exhibit surprising robustness. This capability stems from the property that sub-word embeddings in the embedding space can linearly combine to approximate whole-word embeddings.
Gemstones: A Model Suite for Multi-Faceted Scaling Laws: This work releases the Gemstones model suite — an open-source collection of over 4,000 checkpoints spanning 50M–2B parameters and diverse width-depth ratios. Through systematic experimentation, the paper demonstrates that scaling laws are highly sensitive to design choices such as model selection, learning rate scheduling, and cooldown strategies, and proposes a convex-hull-based fitting method to improve scaling law stability under sparse sampling.
Scalable Fingerprinting of Large Language Models: This paper proposes Perinucleus sampling to generate scalable LLM fingerprints, enabling the embedding of 24,576 fingerprints in Llama-3.1-8B—two orders of magnitude more than existing methods—without degrading model capability. Theoretical and empirical analyses demonstrate that large-scale fingerprinting is essential for defending against collusion attacks.
Alternating Gradient Flows: A Theory of Feature Learning in Two-layer Neural Networks: This paper proposes the Alternating Gradient Flow (AGF) theoretical framework to explain the stepwise "saddle-to-saddle" feature learning dynamics in neural networks. Training is modeled as an alternating process between utility maximization for dormant neurons and cost minimization for active neurons, unifying feature selection analysis across diagonal linear networks, attention models, and modular addition. Predictions from AGF exhibit high agreement with actual gradient flow behavior.
Language Model Behavioral Phases are Consistent Across Architecture, Training Data, and Scale: Through systematic analysis of over 1,400 language model checkpoints—spanning Transformer/Mamba/RWKV architectures, 14M–12B parameter scales, and two training datasets—across 110K+ tokens, this work demonstrates that all autoregressive language models exhibit highly consistent behavioral phases during pre-training: predicted probabilities sequentially overfit to n-gram probabilities of increasing order. Three simple heuristics—word frequency, n-gram probability, and semantic similarity—account for up to 98% of behavioral variance.

Browse all 52 Pretraining papers →

✏️ Knowledge Editing (6)¶

MEMOIR: Lifelong Model Editing with Minimal Overwrite and Informed Retention for LLMs: MEMOIR introduces a framework that incorporates zero-initialized residual memory matrices into FFN layers, employs TopHash-based sparse masks to confine each edit to a distinct subset of memory parameters, and at inference time conditionally activates stored knowledge by measuring mask overlap. The approach achieves an optimal balance among reliability, generalization, and locality across 15,000 sequential edits.
UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models: This paper presents UniEdit — the first unified LLM knowledge editing benchmark built upon an open-domain knowledge graph (Wikidata), covering 311K samples across 25 domains in 5 major categories. By introducing the Neighborhood Multi-hop Chain Sampling (NMCS) algorithm, UniEdit integrates diverse generalization and locality evaluation criteria into a single framework, systematically revealing the shortcomings of existing editing methods under complex ripple effect evaluations.
Rethinking Residual Distribution in Locate-then-Edit Model Editing: This paper reveals that the residual distribution mechanism in locate-then-edit model editing introduces weight deviation errors that grow with distribution distance, batch size, and sequential edit length. It proposes BLUE (Boundary Layer UpdatE), a strategy that updates only the first and last critical layers, achieving an average improvement of 35.59%.
Edit Less, Achieve More: Dynamic Sparse Neuron Masking for Lifelong Knowledge Editing in LLMs: This paper proposes NMKE, a framework that identifies two categories of knowledge neurons—knowledge-general and knowledge-specific—via neuron-level attribution, and applies entropy-guided dynamic sparse masking to achieve precise neuron-level knowledge editing. NMKE maintains high edit success rates and general model capabilities after 5,000 consecutive edits.
KScope: A Framework for Characterizing the Knowledge Status of Language Models: This paper proposes a five-category taxonomy of LLM knowledge status (Consistent Correct / Conflicting Correct / Missing / Conflicting Incorrect / Consistent Incorrect) and the KScope hierarchical statistical testing framework. By combining repeated sampling with multi-step hypothesis testing, KScope precisely characterizes the modal structure of an LLM's knowledge for a given question, and systematically investigates how context updates each knowledge state. The study finds that constrained context summarization combined with credibility augmentation improves knowledge update success rates by an average of 4.3%.
MemEIC: A Step Toward Continual and Compositional Knowledge Editing: This paper proposes MemEIC, a three-tier framework for continual and compositional knowledge editing in large vision-language models (LVLMs), combining an external dual-modal retrieval memory (Mem-E), an internal modality-decoupled LoRA adapter (Mem-I), and a brain-inspired Knowledge Connector. MemEIC substantially outperforms existing methods on the newly introduced CCKEB benchmark.

💬 LLM (Other) (54)¶

Don't Be Lazy: CompleteP Enables Compute-Efficient Deep Transformers: CompleteP parameterization (α=1) is the only scheme that simultaneously achieves hyperparameter transfer along the depth dimension and complete feature learning, saving 12–34% FLOPs over μP on deep models.
Wider or Deeper: Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search: AB-MCTS proposes an adaptive-branching Monte Carlo Tree Search framework that dynamically decides at each node whether to go "wider" (generate new candidate answers) or "deeper" (refine existing answers using feedback), balancing exploration and exploitation via Bayesian posterior updates, and outperforms repeated sampling and standard MCTS on programming and engineering tasks.
Sparse MeZO: Less Parameters for Better Performance in Zeroth-Order LLM Fine-Tuning: This paper proposes Sparse MeZO (S-MeZO), motivated by the observation that zeroth-order gradient noise disproportionately affects parameters with large magnitudes. S-MeZO selectively applies zeroth-order perturbation and updates only to small-magnitude parameters, achieving significant performance gains (+9% on RTE) and convergence acceleration (3.5×) without any additional memory overhead.
StreamBridge: Turning Your Offline Video Large Language Model into a Proactive Streaming Model: StreamBridge proposes a simple and generalizable framework that enables multi-turn streaming interaction via a memory buffer with round-decayed compression, and achieves proactive response through a decoupled lightweight activation model. Combined with the purpose-built Stream-IT dataset, it successfully converts offline Video-LLMs (e.g., Qwen2-VL, LLaVA-OV) into streaming assistants, surpassing GPT-4o and Gemini 1.5 Pro on OVO-Bench and Streaming-Bench.
Do Language Models Use Their Depth Efficiently?: Through causal interventions, residual stream analysis, and cross-model linear mapping, it is demonstrated that the layers in the latter half of current LLMs do not participate in compositional computation but merely iteratively refine the output probability distribution. Deeper models simply stretch the computation of shallower models across more layers.
Detecting High-Stakes Interactions with Activation Probes: Linear activation probes (lightweight classifiers trained on LLM internal representations) are used to detect "high-stakes interactions" from users. Trained on synthetic data, these probes achieve AUROC of 0.88–0.92 across 6 real-world datasets, matching fine-tuned 8–12B LLMs at a computational cost six orders of magnitude lower. A cascaded architecture (probe pre-filtering + LLM refinement) further surpasses either component used alone.
AutoDiscovery: Open-ended Scientific Discovery via Bayesian Surprise: AutoDiscovery proposes Bayesian Surprise as an objective reward signal for open-ended scientific discovery — estimating the KL divergence between prior and posterior belief distributions via LLM sampling, combined with MCTS and progressive widening to explore the hypothesis space. On 21 real-world datasets, the method produces 5–29% more surprising discoveries than greedy/beam search baselines. Human evaluation confirms that Bayesian Surprise aligns with expert "surprise" ratings (0.67), substantially outperforming LLM self-evaluated "novelty" and "usefulness."
Solving Inequality Proofs with Large Language Models: This paper proposes IneqMath, the first large-scale olympiad-level inequality benchmark, formulates inequality proving as two automatically verifiable subtasks (bound estimation and relation prediction), develops a five-module LLM-as-Judge framework, and finds that even o1 achieves an overall accuracy below 10% under step-by-step reasoning scrutiny.
Breaking AR's Sampling Bottleneck: Provable Acceleration via Diffusion Language Models: This paper establishes a complete convergence theory for masked diffusion language models from an information-theoretic perspective: it proves that the sampling error in KL divergence decays at an $O(1/T)$ rate and scales linearly with inter-token mutual information, provides a matching lower bound to establish tightness, and theoretically demonstrates that diffusion models can generate high-quality samples in $T < L$ steps (where $L$ is the sequence length).
System Prompt Optimization with Meta-Learning: This paper formulates system prompt optimization as a bilevel problem and proposes MetaSPO, a meta-learning framework that optimizes system prompts for cross-task generalization in the outer loop while optimizing task-specific user prompts in the inner loop. The resulting system prompts significantly outperform baselines across 14 unseen tasks.

Browse all 54 LLM (Other) papers →

📖 NLP Understanding (3)¶

Planning without Search: Refining Frontier LLMs with Offline Goal-Conditioned RL: This paper proposes PNLC, a method that trains a lightweight goal-conditioned value function as a "natural language critic" to guide LLM agents in multi-turn planning and self-refinement at the thought-step level. Without direct fine-tuning or inference-time search, PNLC significantly outperforms existing methods on complex interactive tasks such as web navigation, social reasoning, and persuasion, while achieving 8–10× faster inference.
Generalization Error Analysis for Selective State-Space Models Through the Lens of Attention: This work unrolls selective SSMs (Mamba) into an attention-equivalent form and derives generalization bounds via covering number techniques, controlled by the spectral abscissa $s_{\mathbf{A}}$ of the continuous-time state matrix. When $s_{\mathbf{A}} < 0$, the bound is independent of sequence length; when $s_{\mathbf{A}} \geq 0$, it grows exponentially. The paper further proves this dependence is irreducible.
Weak-to-Strong Generalization under Distribution Shifts: This paper demonstrates that naive weak-to-strong generalization fails under distribution shifts—where the strong model performs even worse than the weak supervisor—and proposes RAVEN, a framework that dynamically learns optimal combination weights over multiple weak models to achieve robust weak-to-strong generalization, surpassing baselines by over 30% on OOD tasks.

🗣️ Dialogue Systems (8)¶

SciArena: An Open Evaluation Platform for Non-Verifiable Scientific Literature-Grounded Tasks: SciArena is a community-driven open evaluation platform for scientific literature tasks. It adopts a Chatbot Arena-style human preference voting paradigm to rank 47 foundation models, collecting over 20,000 votes, and releases SciArena-Eval as a meta-benchmark for assessing the ability of automated evaluation systems to judge answer quality on literature-grounded tasks.
HyGen: Efficient LLM Serving via Elastic Online-Offline Request Co-location: This paper proposes HyGen, an interference-aware LLM inference system that achieves elastic co-location of online and offline workloads through an accurate batch latency predictor, an SLO-aware performance profiler, and a prefix-sharing-maximization scheduling strategy, delivering 3.87–5.84× throughput gains while strictly guaranteeing SLO compliance.
Bridging Human and LLM Judgments: Understanding and Narrowing the Gap: This paper proposes Bridge, a statistical framework that models the latent relationship between human and LLM judgments via ordinal logistic regression. With a small number of human labels, Bridge improves the calibration and alignment of LLM judgments while supporting formal statistical hypothesis testing for systematic biases.
KL Penalty Control via Perturbation for Direct Preference Optimization: This paper proposes ε-DPO, which achieves instance-level adaptive KL penalty control by monitoring the monotonicity of logits—used as preference model outputs—under small perturbations of $\beta$ during training. The method incurs no additional computational overhead and significantly outperforms DPO and most direct alignment algorithms, achieving a 46.4% LC win rate on AlpacaEval 2 (vs. 40.3% for DPO).
AC-LoRA: (Almost) Training-Free Access Control-Aware Multi-Modal LLMs: AC-LoRA is an end-to-end system that trains independent LoRA adapters for datasets with different permission levels. At inference time, it dynamically retrieves and training-freely merges multiple LoRA outputs based on cosine similarity and user permissions, achieving strong information isolation while matching or surpassing SOTA LoRA mixture methods in response quality.
LatentGuard: Controllable Latent Steering for Robust Refusal of Attacks and Reliable Response Generation: This paper proposes LatentGuard, a three-stage framework that combines behavior-level alignment fine-tuning, structured VAE-supervised latent space modeling, and latent-space dimensional manipulation to achieve interpretable and controllable regulation of LLM refusal behavior — robustly defending against adversarial attacks while preserving responsiveness to benign queries.
Agentic Persona Control and Task State Tracking for Realistic User Simulation: A three-agent collaborative framework for realistic user simulation is proposed, comprising a User Agent (coordination), a State Tracking Agent (structured task state), and a Message Attributes Generation Agent (behavior attribute control conditioned on persona and state). On a restaurant ordering scenario, the framework achieves a 102.6% improvement in composite realism score (CRRS), +19.9% in persona adherence, and +284.5% in behavioral variability. A core finding is that behavior control without state awareness yields BVS = 0 (completely rigid behavior).
Less is More: Local Intrinsic Dimensions of Contextual Language Models: This paper proposes using the Local Intrinsic Dimension (LID) of contextual token embeddings as an unsupervised signal for monitoring LLM training dynamics — a decrease in LID indicates improved generalization, while an increase signals overfitting. The utility of this geometric signal is validated on tasks including dialogue state tracking, grokking, and sentiment recognition.

🌐 Multilingual & Translation (11)¶

HelpSteer3-Preference: Open Human-Annotated Preference Data across Diverse Tasks and Languages: NVIDIA releases a 40K+ open-source human-annotated preference dataset covering general, STEM, code, and multilingual (13 languages) tasks. The reward model trained on this dataset achieves 82.4% (+10%) on RM-Bench, with a commercially friendly CC-BY-4.0 license.
XIFBench: Evaluating Large Language Models on Multilingual Instruction Following: This paper proposes XIFBench — the first constraint-driven benchmark systematically evaluating LLMs' multilingual instruction-following capabilities. It comprises 558 instructions (0–5 constraints, 5 categories × 21 dimensions) across 6 languages (high/mid/low resource), and introduces an English-requirement anchoring evaluation protocol that achieves 94.7% cross-lingual evaluation consistency.
On Extending Direct Preference Optimization to Accommodate Ties: This paper replaces the Bradley-Terry preference model in DPO with the Rao-Kupper and Davidson extensions, enabling preference optimization to explicitly model "tie" data. This avoids discarding ambiguous preference pairs and yields improved regularization and performance on translation and mathematical reasoning tasks.
MERIT: Multilingual Semantic Retrieval with Interleaved Multi-Condition Query: This paper introduces MERIT, the first multilingual interleaved multi-condition semantic retrieval dataset (320K queries, 135K products, 5 languages, 7 product categories), exposes the bottleneck of existing retrieval models that focus solely on global semantics while neglecting condition-level details, and proposes the Coral fine-tuning framework that combines embedding reconstruction with contrastive learning to achieve a 45.9% improvement in retrieval performance.
Exploring the Translation Mechanism of Large Language Models: This paper proposes a subspace-intervened path patching method for fine-grained causal analysis of the translation mechanism in LLMs. The study finds that translation is driven by a sparse set of attention heads comprising fewer than 5% of all heads, categorized into three functional roles: source heads, indicator heads, and positional heads. MLP layers integrate these features into an English-centric intermediate representation, and fine-tuning only 64 critical heads achieves performance comparable to full-parameter fine-tuning.
ParallelPrompt: Extracting Parallelism from Large Language Model Queries: This work presents ParallelPrompt, the first benchmark for intra-query parallelism, comprising structured decomposition annotations for 37,000+ real user prompts. It demonstrates that approximately 10% of user queries contain exploitable parallel structure, and that parallel execution can achieve up to 5.7× latency speedup with limited quality degradation.
Zero-Shot Performance Prediction for Probabilistic Scaling Laws: This paper frames NLP learning curve prediction as a multi-task learning problem, employing a latent-variable multi-output Gaussian process (MaGP) to capture the bi-level hierarchical structure of datasets and inter-task correlations, enabling zero-shot prediction of learning curves and deriving probabilistic scaling laws via Monte Carlo simulation.
How Data Mixing Shapes In-Context Learning: Asymptotic Equivalence for Transformers with MLPs: Under a high-dimensional asymptotic framework, this paper proves that Transformers with nonlinear MLP heads are asymptotically equivalent to structured polynomial predictors in terms of ICL error, revealing the gain mechanism of nonlinear MLPs on nonlinear tasks and establishing that low noise and structured covariance are key characteristics of high-quality data sources in multi-source data mixing.
Adaptive Originality Filtering: Rejection-Based Prompting and RiddleScore for Culturally Grounded Multilingual Riddle Generation: This paper proposes Adaptive Originality Filtering (AOF)—a semantic rejection-sampling prompting strategy that filters repetitive or templated outputs via MiniLM embedding cosine similarity, compelling LLMs to generate more novel, diverse, and culturally grounded multilingual riddles. It also introduces the RiddleScore composite evaluation metric (Novelty + Diversity + Fluency + Alignment), achieving a human correlation of $\rho=0.83$.
Quantifying Climate Policy Action and Its Links to Development Outcomes: A Cross-National Data-Driven Analysis: This paper constructs an integrated NLP–econometrics framework that first uses a fine-tuned multilingual DistilBERT to automatically classify global climate policy documents by topic (Mitigation / Adaptation / Disaster Risk Management / Loss & Damage) with F1 = 0.90, then conducts fixed-effects panel regression against World Bank development indicators, finding that mitigation policies are significantly positively associated with higher GDP/GNI, while Loss & Damage policies remain substantially unimplemented worldwide.

Browse all 11 Multilingual & Translation papers →

🔍 Information Retrieval & RAG (25)¶

HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation: This paper proposes HyperGraphRAG, the first RAG method based on hypergraph structure, which models n-ary relations ($n \geq 2$) via hyperedges. It overcomes the binary-relation bottleneck of existing graph-based RAG methods, achieving comprehensive improvements over StandardRAG and the GraphRAG family on question-answering tasks across medical, agricultural, computer science, and legal domains.
Chain-of-Retrieval Augmented Generation (CoRAG): This paper proposes CoRAG, a framework that automatically generates intermediate retrieval chains (sub-query → sub-answer) via rejection sampling, fine-tunes an LLM to learn iterative retrieval and reasoning, and supports diverse test-time decoding strategies (greedy / Best-of-N / tree search) for flexible compute scaling. CoRAG achieves 26+ EM improvement on multi-hop QA and attains state-of-the-art on 9/10 tasks of the KILT benchmark.
Scaling Language-Centric Omnimodal Representation Learning: This paper proposes the LCO-Emb framework and demonstrates that Multimodal Large Language Models (MLLMs) implicitly establish cross-modal alignment during generative pretraining. Lightweight text-only contrastive fine-tuning suffices to activate full omnimodal representation capabilities. The work further identifies the Generation-Representation Scaling Law (GRSL), which establishes a positive correlation between generative capability and representation performance.
SymRTLO: Enhancing RTL Code Optimization with LLMs and Neuron-Inspired Symbolic Reasoning: This paper proposes SymRTLO, the first neurosymbolic framework integrating LLMs with symbolic reasoning for RTL code optimization. By combining retrieval-augmented optimization rules, AST template-guided code generation, and an FSM symbolic system, SymRTLO achieves improvements of up to 43.9%, 62.5%, and 51.1% in power, performance, and area (PPA), respectively.
Is PRM Necessary? Problem-Solving RL Implicitly Induces PRM Capability in LLMs: This study systematically demonstrates that pure RL training (without explicit PRM supervision) implicitly induces strong process judgment capability; existing PRMs are even less effective than simple majority voting on strong reasoning models such as DeepSeek-R1 and QwQ-32B. The paper proposes Self-PRM, which allows a model to rerank its outputs using its own internal reward signal, consistently outperforming external PRMs.
AcuRank: Uncertainty-Aware Adaptive Computation for Reranking: AcuRank dynamically adjusts the reranking subset size and verification scope via TrueSkill-based uncertainty estimation, achieving a superior accuracy–efficiency trade-off while avoiding over-computation.
Learning Task-Agnostic Representations through Multi-Teacher Distillation: This paper proposes a task-agnostic multi-teacher distillation framework based on mutual information maximization. By estimating the conditional distribution of teacher embeddings via Gaussian kernels, the student model learns high-information-density general-purpose representations without relying on any downstream task labels, achieving state-of-the-art performance among same-scale models across text, vision, and molecular modeling domains.
SuperCLIP: CLIP with Simple Classification Supervision: SuperCLIP augments the CLIP contrastive learning framework with an extremely simple classification loss — requiring only a lightweight linear layer that increases total FLOPs by merely 0.077% — to recover fine-grained textual supervision that CLIP underutilizes, achieving consistent improvements on zero-shot classification, image-text retrieval, and vision-only tasks.
Hierarchical Retrieval: The Geometry and a Pretrain-Finetune Recipe: This paper investigates the feasibility of Dual Encoders (DE) for Hierarchical Retrieval (HR), theoretically proving that embedding dimensionality need only grow linearly with hierarchy depth and logarithmically with document count. After identifying the "lost-in-the-long-distance" phenomenon, the paper proposes a pretrain-finetune strategy that improves long-distance recall from 19% to 76% on WordNet.
MuRating: A High Quality Data Selecting Approach to Multilingual Large Language Model Pretraining: This paper proposes MuRating, a scalable multilingual data selection framework that aggregates multiple English data quality scorers via pairwise comparisons, transfers quality signals to 17 languages through translation, and trains a language-agnostic multilingual quality assessment model, achieving consistent performance gains in LLM pretraining at both 1.2B and 7B scales.

Browse all 25 Information Retrieval & RAG papers →

💻 Code Intelligence (19)¶

SWE-rebench: An Automated Pipeline for Task Collection and Decontaminated Evaluation of Software Engineering Agents: A fully automated pipeline is developed to continuously mine real-world software engineering interaction tasks from GitHub, producing the SWE-rebench dataset of 21,000+ executable Python tasks and a decontaminated benchmark. The work reveals that several models exhibit contamination-inflated performance on SWE-bench Verified (e.g., DeepSeek-V3: 39.7% on SWE-bench vs. 21.3% on SWE-rebench).
MLR-Bench: Evaluating AI Agents on Open-Ended Machine Learning Research: This paper proposes MLR-Bench, a comprehensive benchmark comprising 201 open-ended ML research tasks, accompanied by MLR-Judge (an LLM-based evaluation framework) and MLR-Agent (a modular research agent). The study finds that state-of-the-art coding agents fabricate or fail to verify experimental results in approximately 80% of cases, exposing a fundamental bottleneck in AI-automated scientific research.
A Self-Improving Coding Agent: This paper proposes SICA (Self-Improving Coding Agent), a coding agent capable of autonomously editing its own codebase to improve performance. By eliminating the distinction between meta-agent and target-agent, SICA achieves iterative self-improvement, advancing from 17% to 53% on a subset of SWE-Bench Verified.
Co-Evolving LLM Coder and Unit Tester via Reinforcement Learning: This paper proposes CURE, a framework in which a single LLM simultaneously assumes the roles of code generator and unit test generator. Cross-execution between generated code and generated tests constructs a pairwise reward matrix; theoretically derived reward signals then drive reinforcement learning. Without any ground-truth code annotations, CURE achieves co-evolution of both code generation and unit test generation capabilities, substantially outperforming dedicated coder models of comparable scale across five programming benchmarks.
MaintainCoder: Maintainable Code Generation Under Dynamic Requirements: This work is the first to systematically define and address the maintainability problem in LLM-based code generation, contributing both a benchmark and a method. MaintainBench evaluates code maintainability under requirement evolution using 4 change patterns and dynamic metrics; MaintainCoder integrates the Waterfall model, design patterns, and 6 specialized agents, achieving 60%+ improvement on dynamic maintainability metrics while also improving initial code correctness.
CodeCrash: Exposing LLM Fragility to Misleading Natural Language in Code Reasoning: This paper proposes CodeCrash, a stress-testing framework that systematically evaluates the code reasoning robustness of 17 LLMs through functionally equivalent structural perturbations and misleading natural language injections (comments, print statements, and hints). The framework reveals an average performance drop of 23.2% across models, with CoT recovering only to 13.8%, and is the first to identify the "Reasoning Collapse" phenomenon in large reasoning models (LRMs).
AstroVisBench: A Code Benchmark for Scientific Computing and Visualization in Astronomy: AstroVisBench introduces the first code benchmark for evaluating LLMs on astronomical scientific computing and visualization. It extracts 864 tasks (processing + visualization) from 110 Jupyter Notebooks, and designs a dual evaluation pipeline (execution-based variable inspection + VLM-as-Judge visualization scoring, achieving Spearman ρ=0.822 with expert ratings). Evaluation of 8 state-of-the-art models reveals that Gemini 2.5 Pro performs best, yet attains only a 15.7% error-free rate, with FileNotFoundError accounting for 43% of all errors.
Preserving LLM Capabilities through Calibration Data Curation: From Analysis to Optimization: This paper systematically investigates how compositional properties of calibration data (sequence length, sample size, source, format) and domain correspondence affect capability preservation after LLM compression. It finds that representativeness and diversity in the activation space are the fundamental determinants of calibration data quality, and proposes a three-stage calibration data curation framework, COLA.
QiMeng-SALV: Signal-Aware Learning for Verilog Code Generation: This paper proposes QiMeng-SALV, a signal-aware learning method that extracts functionally correct signal-level code snippets from partially incorrect Verilog modules as reward signals for DPO training, elevating the optimization granularity from module level to signal level and achieving SOTA on VerilogEval and RTLLM.
Program Synthesis via Test-Time Transduction: This paper proposes SYNTRA, a framework that reframes program synthesis as transductive learning — at test time, it leverages visible test inputs and LLM judgment to iteratively eliminate inconsistent candidate program hypotheses. A greedy maximin algorithm minimizes the number of LLM queries, achieving accuracy improvements of up to 196% across 4 benchmarks.

Browse all 19 Code Intelligence papers →

🎨 Image Generation (222)¶

Real-Time Execution of Action Chunking Flow Policies: This paper proposes Real-Time Chunking (RTC), which frames asynchronous action chunk execution as an inpainting problem. By freezing already-executed actions and "inpainting" the remainder to be consistent with the prefix, RTC enables smooth real-time execution of diffusion/flow policies without any retraining.
Show-o2: Improved Native Unified Multimodal Models: This paper presents Show-o2, a natively unified multimodal model built upon autoregressive modeling and Flow Matching. By constructing unified visual representations in a 3D causal VAE space via dual-path spatial(-temporal) fusion, Show-o2 supports multimodal understanding and generation across text, images, and video, with a two-stage training strategy that effectively preserves language knowledge.
Why Diffusion Models Don't Memorize: The Role of Implicit Dynamical Regularization in Training: Through numerical experiments and theoretical analysis, this paper identifies two critical timescales in diffusion model training — a generalization time $\tau_{\text{gen}}$ and a memorization time $\tau_{\text{mem}}$ — where the latter scales linearly with training set size $n$ while the former remains constant. The resulting implicit dynamical regularization enables early stopping to prevent memorization even in heavily overparameterized regimes.
Dual Data Alignment Makes AI-Generated Image Detector Easier Generalizable: This paper proposes Dual Data Alignment (DDA), which generates synthetic training images via pixel-domain and frequency-domain dual alignment to eliminate spurious correlations caused by dataset bias. By forcing the detector to learn only forgery-relevant features, DDA achieves an average accuracy of 90.7% across 11 benchmarks, substantially outperforming existing methods.
Is Artificial Intelligence Generated Image Detection a Solved Problem?: This paper proposes AIGIBench, a comprehensive benchmark that systematically evaluates 11 state-of-the-art detectors across four tasks—multi-source generalization, multi-degradation robustness, data augmentation sensitivity, and test-time preprocessing impact—revealing that existing AIGI detection methods suffer severe performance degradation in real-world scenarios, demonstrating that the problem is far from solved.
RLVR-World: Training World Models with Reinforcement Learning: This paper proposes the RLVR-World framework, extending the Reinforcement Learning with Verifiable Rewards (RLVR) paradigm to world model training. By directly optimizing target metrics (e.g., prediction accuracy, perceptual quality) as verifiable rewards, the framework achieves significant improvements on both language and video world models.
Fast Solvers for Discrete Diffusion Models: Theory and Applications of High-Order Algorithms: This work introduces high-order numerical methods into discrete diffusion model inference for the first time, proposing two second-order solvers — θ-RK-2 and θ-Trapezoidal — and theoretically proving that θ-Trapezoidal improves the discretization error from first-order $\mathcal{O}(\kappa T)$ to second-order $\mathcal{O}(\kappa^2 T)$. Experiments spanning 200M–8B models consistently demonstrate improvements across text, image, and mathematical reasoning tasks.
Boosting Generative Image Modeling via Joint Image-Feature Synthesis: This paper proposes ReDi (Representation Diffusion), a framework that jointly models VAE image latents and DINOv2 semantic features within a diffusion model — both are simultaneously denoised from pure noise within a single diffusion process. With minimal modifications to the DiT architecture, ReDi achieves a 23× training convergence speedup and state-of-the-art FID, while unlocking a novel Representation Guidance inference strategy.
Failure Prediction at Runtime for Generative Robot Policies: This paper proposes FIPER, a framework for runtime failure prediction in generative robot policies (diffusion/flow matching). It jointly evaluates an observation-side metric RND-OE (OOD detection) and an action-side metric ACE (Action Chunk Entropy) to enable early and accurate failure prediction without any failure data, with statistical guarantees provided via conformal prediction.
Inference-Time Scaling for Flow Models via Stochastic Generation and Rollover Budget Forcing: This paper proposes inference-time scaling methods for flow models: stochasticity is introduced via ODE→SDE conversion to enable particle sampling; the search space is expanded through linear→VP interpolant conversion; and a Rollover Budget Forcing (RBF) strategy is designed to adaptively allocate the computational budget. The approach substantially outperforms all existing methods on compositional text-to-image generation and quantity-aware generation tasks.

Browse all 222 Image Generation papers →

🎬 Video Generation (23)¶

Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion: This paper proposes the Self Forcing training paradigm, which eliminates the exposure bias caused by train-inference distribution mismatch in Teacher Forcing and Diffusion Forcing by performing autoregressive self-rollout during training and applying a holistic video-level distribution matching loss (DMD/SiD/GAN). Built on Wan2.1-1.3B, it achieves real-time streaming video generation at 17 FPS on a single GPU while matching or surpassing the quality of bidirectional diffusion models that are orders of magnitude slower.
Scaling RL to Long Videos: This paper proposes LongVILA-R1, a full-stack framework that extends VLM reasoning to long videos (up to 8192 frames) via a 104K long-video reasoning dataset, a two-stage CoT-SFT + RL training pipeline, and the MR-SP multimodal reinforcement sequence parallelism system, achieving 65.1%/71.1% on VideoMME.
Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation: This paper proposes AAPT (Autoregressive Adversarial Post-Training), which converts a pretrained video diffusion model into an autoregressive real-time video generator via adversarial training. The model requires only one forward pass per frame (1NFE), employs student-forcing training to reduce error accumulation, and achieves real-time streaming generation at 736×416 resolution and 24fps on a single H100 GPU, supporting videos up to one minute in length (1440 frames).
Force Prompting: Video Generation Models Can Learn and Generalize Physics-based Control Signals: This paper proposes Force Prompting, which uses physical forces (local point forces and global wind forces) as control signals for video generation models. Using only ~15K synthetic training videos (Blender flags and rolling balls) and a single day of training on 4×A100 GPUs, the method achieves remarkable generalization across diverse real-world scenes with varying objects, materials, and geometries, including preliminary mass understanding capabilities.
PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation: PhysCtrl employs diffusion models to learn the physical dynamics distribution of four material types (elastic, sand, plasticine, and rigid bodies), representing dynamics as 3D point trajectories. A diffusion model incorporating spatiotemporal attention and physics constraints is trained on 550K synthetic animations; the generated trajectories drive a pretrained video model to achieve high-fidelity physics video generation controllable by force and material parameters.
MagCache: Fast Video Generation with Magnitude-Aware Cache: This paper discovers that the magnitude ratio of residual outputs between adjacent timesteps in video diffusion models follows a universally monotonically decreasing pattern across models and prompts — termed the "Unified Magnitude Law" — and proposes MagCache: a method that accurately models skip-step error accumulation via magnitude ratios, adaptively skips redundant timesteps and reuses cached outputs with only a single calibration sample, achieving 2.10–2.68× speedup on Open-Sora, CogVideoX, Wan 2.1, and HunyuanVideo while outperforming TeaCache and other existing methods across all three metrics of LPIPS, SSIM, and PSNR.
Training-Free Efficient Video Generation via Dynamic Token Carving: This paper proposes Jenga, a training-free inference acceleration framework for video DiTs that achieves 8.83× speedup on HunyuanVideo with only a 0.01% drop in VBench score. The framework combines dynamic block attention carving (sparse KV block selection after token reordering via 3D space-filling curves) and a progressive resolution strategy (coarse-to-fine denoising), which operate orthogonally.
DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models: This paper identifies and addresses the motion bias problem in video DPO — by constructing structurally aligned video pairs via noising and denoising GT videos to fix the motion dimension, annotating dense preferences at the temporal segment level for more precise learning signals, and leveraging off-the-shelf VLMs for automatic annotation to reduce cost. Using only 1/3 of the annotation data, the method substantially improves motion generation quality while matching visual quality and text alignment.
VORTA: Efficient Video Diffusion via Routing Sparse Attention: This paper proposes VORTA, a framework that achieves end-to-end 1.76× acceleration of video diffusion Transformers without quality degradation, through bucketed coreset attention (for modeling long-range dependencies) and a signal-aware routing mechanism (for adaptively selecting sparse attention branches). Combined with caching and distillation methods, it achieves up to 14.41× acceleration.
S²Q-VDiT: Accurate Quantized Video Diffusion Transformer with Salient Data and Sparse Token Distillation: To address the high calibration variance and optimization difficulty caused by extremely long token sequences in video diffusion Transformers, this paper proposes the S²Q-VDiT framework. By combining Hessian-aware salient data selection and attention-guided sparse token distillation, it achieves lossless quantization under W4A6 settings for the first time, yielding 3.9× model compression and 1.3× inference speedup.

Browse all 23 Video Generation papers →

🧩 Multimodal VLM (108)¶

UniTok: A Unified Tokenizer for Visual Generation and Understanding: This paper proposes UniTok, a unified tokenizer for visual generation and understanding that overcomes the representation capacity bottleneck of discrete tokens via Multi-Codebook Quantization (MCQ). UniTok achieves simultaneous state-of-the-art records of 0.38 rFID and 78.6% zero-shot accuracy on ImageNet, and can be seamlessly integrated into MLLMs to enable both generation and understanding.
ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation: This paper proposes ForceVLA, which introduces 6-axis force/torque sensing as a first-class modality within the VLA framework. A Force-aware Vision-Language Mixture-of-Experts (FVLMoE) module dynamically fuses visual-language embeddings with real-time force feedback at the action decoding stage, achieving an average success rate improvement of 23.2% across five contact-rich manipulation tasks, with individual tasks reaching up to 80%.
Learning from Videos for 3D World: Enhancing MLLMs with 3D Vision Geometry Priors: VG-LLM proposes integrating a 3D visual geometry encoder (VGGT) into multimodal large language models, enabling the extraction and fusion of 3D geometric priors from video input alone—without any explicit 3D data. This approach significantly improves MLLM performance on 3D scene understanding and spatial reasoning tasks, with the 4B model surpassing Gemini-1.5-Pro on VSI-Bench.
What Can RL Bring to VLA Generalization? An Empirical Study: This paper systematically investigates the effect of RL fine-tuning on the generalization capabilities of Vision-Language-Action (VLA) models. The study finds that PPO is the most effective RL algorithm, significantly outperforming DPO and GRPO; RL yields substantially greater OOD generalization than SFT in semantic understanding and execution robustness, while achieving comparable visual robustness.
BridgeVLA: Input-Output Alignment for Efficient 3D Manipulation Learning with Vision-Language Models: This paper proposes BridgeVLA, which projects 3D point clouds into multi-view 2D images and uses 2D heatmaps as an intermediate representation to align the input and output spaces, enabling efficient and effective 3D robot manipulation learning.
Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models: This paper extends sparse autoencoders (SAEs) to vision-language models (e.g., CLIP), proposes the MonoSemanticity score (MS) to quantitatively evaluate the monosemanticity of neurons, and demonstrates that manipulating SAE neurons can directly steer multimodal large language models (e.g., LLaVA) to insert or suppress specific concepts.
BioCLIP 2: Emergent Properties from Scaling Hierarchical Contrastive Learning: BioCLIP 2 trains a ViT-L on TreeOfLife-200M (214M images across 952K species) using hierarchical contrastive learning, achieving an 18% improvement over BioCLIP in zero-shot species recognition. The work further uncovers emergent properties arising from scale: embeddings automatically encode ecological relationships (e.g., Darwin's finches arranged by beak size), and intra-species variation is orthogonal to inter-species variation.
GEM: Empowering MLLM for Grounded ECG Understanding with Time Series and Images: GEM is proposed as the first multimodal large language model that unifies ECG time series, 12-lead ECG images, and text. Through a dual-encoder framework, cross-modal alignment, and knowledge-guided instruction data generation, GEM achieves grounded ECG diagnosis based on quantifiable physiological features, improving diagnostic accuracy by 7.4%, interpretability by 22.7%, and grounding capability by 25.3%.
Benchmarking Retrieval-Augmented Multimodal Generation for Document Question Answering: This paper introduces the MMDocRAG benchmark (4,055 expert-annotated QA pairs) to systematically evaluate 60 VLMs/LLMs and 14 retrievers on quote selection and interleaved text-image answer generation in multimodal document RAG. Results reveal that the strongest model, GPT-4.1, achieves only 70.2% Quote Selection F1, while fine-tuning yields substantial performance gains.
Don't Just Chase "Highlighted Tokens" in MLLMs: Revisiting Visual Holistic Context Retention: This paper proposes HoloV, a plug-and-play visual token pruning framework that adaptively allocates pruning budgets across different spatial crop regions to preserve global visual context rather than retaining only attention-highlighted tokens. On LLaVA-1.5, HoloV retains 95.8% of original performance after pruning 88.9% of visual tokens.

Browse all 108 Multimodal VLM papers →

🧠 VLM Reasoning (30)¶

Video-R1: Reinforcing Video Reasoning in MLLMs: Inspired by DeepSeek-R1, this paper presents the first systematic exploration of applying the R1 paradigm (rule-based RL) to video reasoning. It proposes the T-GRPO algorithm to explicitly encourage temporal reasoning, constructs a mixed image-video training dataset, and achieves 37.1% accuracy on VSI-Bench, surpassing GPT-4o.
RoboRefer: Towards Spatial Referring with Reasoning in Vision-Language Models for Robotics: This paper proposes RoboRefer, a 3D-aware reasoning VLM trained via a two-stage SFT + RFT strategy with a metric-sensitive process reward function. It achieves precise single-step spatial understanding and multi-step spatial reasoning on spatial referring tasks, surpassing Gemini-2.5-Pro by 17.4% on RefSpatial-Bench.
VAGEN: Reinforcing World Model Reasoning for Multi-Turn VLM Agents: VAGEN is a framework that structures the reasoning process of VLM agents into StateEstimation and TransitionModeling to build an internal world model, and combines WorldModeling Reward with Bi-Level GAE for efficient multi-turn RL training. A 3B model trained under this framework (0.82) surpasses GPT-5 (0.75) and Gemini 2.5 Pro (0.67).
SSR: Enhancing Depth Perception in VLMs via Rationale-Guided Spatial Reasoning: This paper proposes the SSR framework, which converts raw depth information into structured textual reasoning rationales and compresses them into compact latent embeddings via knowledge distillation, enhancing the spatial reasoning capabilities of existing VLMs in a plug-and-play manner.
Recognition through Reasoning: Reinforcing Image Geo-localization with Large Vision-Language Models: This paper proposes GLOBE — an LVLM-based image geo-localization system trained via GRPO reinforcement learning. By constructing MP16-Reason, a reasoning-oriented dataset with localizability assessment, visual-clue reasoning chains, and geographic accuracy annotations, GLOBE surpasses SOTA methods trained on millions of samples as well as large-scale open-source VLMs using only 33K training examples across multiple benchmarks.
RTV-Bench: Benchmarking MLLM Continuous Perception, Understanding and Reasoning through Real-Time Video: This paper proposes RTV-Bench, a fine-grained evaluation benchmark for assessing the continuous real-time video analysis capabilities of MLLMs. Comprising 552 videos and 4,608 QA pairs, it comprehensively evaluates model perception, understanding, and reasoning in dynamic video streams through a multi-timestamp QA mechanism, hierarchical question structure, and multi-dimensional assessment.
AffordBot: 3D Fine-grained Embodied Reasoning via Multimodal Large Language Models: This paper introduces a fine-grained 3D embodied reasoning task—jointly predicting the spatial location, motion type, and motion axis of actionable elements—and proposes rendering 3D point clouds into panoramic views with projected affordance candidates, guided by a customized Chain-of-Thought (CoT) reasoning paradigm for MLLMs, achieving state-of-the-art performance with AP25 of 23.3%.
MIRAGE: A Benchmark for Multimodal Information-Seeking and Reasoning in Agriculture: MIRAGE is the first multimodal benchmark constructed from real agricultural expert consultation dialogues (35,000+), evaluating vision-language models on domain-level entity identification, causal reasoning, and clarify-or-respond decision-making. It reveals a severe challenge in which even GPT-4.1 achieves only 43.9% identification accuracy.
Sherlock: Self-Correcting Reasoning in Vision-Language Models: The first systematic study of self-correction capabilities in reasoning VLMs: existing reasoning VLMs are found to be nearly incapable of self-correction (<10% exhibit an aha moment). The paper proposes Sherlock, a three-stage training framework (SFT cold-start → offline trajectory-level preference learning → online self-iterative improvement) that surpasses LLaVA-CoT/Mulberry/LlamaV-o1 (which use 100K–260K annotations) using only 20K labeled samples.
Retrv-R1: A Reasoning-Driven MLLM Framework for Universal and Efficient Multimodal Retrieval: This paper proposes Retrv-R1, the first R1-style reasoning-based multimodal retrieval framework. It reduces token consumption via an Information Compression Module (ICM), preserves complete information for hard candidates through a Details Inspection Mechanism (DIM), and employs a curriculum-based RL reward to balance effectiveness and efficiency, achieving state-of-the-art performance on universal multimodal retrieval benchmarks.

Browse all 30 VLM Reasoning papers →

⚡ VLM Efficiency (8)¶

FlowCut: Rethinking Redundancy via Information Flow for Efficient Vision-Language Models: Reexamining visual token redundancy in VLMs through the lens of Information Flow: the CLS token acts as an information relay, redundancy emerges progressively, and single-layer single-criterion scoring is unreliable. FlowCut—an information-flow-aware multi-criteria cumulative importance pruning framework—surpasses SOTA by 1.6% on LLaVA-1.5-7B at an 88.9% token reduction rate, and by 4.3% on LLaVA-NeXT-7B.
Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization: This paper proposes Balanced Token Pruning (BTP), which jointly considers the impact of pruning on both the current layer (local) and subsequent layers (global). BTP emphasizes diversity preservation in shallow layers to maintain downstream representation quality, and attention-based selection in deep layers to preserve local output consistency. On multiple LVLMs including LLaVA and Qwen2.5-VL, BTP retains 98% of the original model's performance while keeping only 22% of visual tokens.
ViSpec: Accelerating Vision-Language Models with Vision-Aware Speculative Decoding: To address the difficulty of draft models in handling redundant visual tokens during VLM speculative decoding, this paper proposes ViSpec, a framework that achieves significant acceleration (up to 3.22×) in VLM speculative decoding for the first time, via a visual adapter for image token compression, global visual feature injection, and synthetic training data generation.
SCOPE: Saliency-Coverage Oriented Token Pruning for Efficient Multimodal LLMs: This paper proposes SCOPE, a visual token pruning strategy that jointly models saliency and coverage. By iteratively selecting tokens with the highest SCOPE scores, it preserves semantic completeness and retains 96% of LLaVA-1.5's performance under a 9× token reduction.
PrefixKV: Adaptive Prefix KV Cache is What Vision Instruction-Following Models Need for Efficient Generation: PrefixKV identifies that the importance distributions of KV caches vary substantially across layers, and formalizes the per-layer cache sizing problem as a global prefix configuration search. A binary search is employed to find the optimal cumulative priority threshold that maximizes contextual information retention in each layer. At a 20% retention ratio, PrefixKV incurs only a 0.49 PPL degradation while delivering a 1.8× inference speedup.
ElasticMM: Efficient MLLM Serving with Elastic Multimodal Parallelism: This paper proposes the Elastic Multimodal Parallelism (EMP) paradigm and the ElasticMM system, which disaggregates different stages of multimodal inference into independent instances via modality-aware load balancing and elastic partition scheduling, achieving up to 4.2× TTFT reduction and 3.2–4.5× throughput improvement over vLLM.
HAWAII: Hierarchical Visual Knowledge Transfer for Efficient VLM: This paper proposes the Hawaii framework, which distills knowledge from multiple visual experts into a single visual encoder via Mixture of LoRA Adapters (MoLA) and Hierarchical Knowledge Distillation (HKD), significantly improving the visual understanding capability of VLMs without incurring any additional inference cost.
Beyond Greedy Exits: Improved Early Exit Decisions for Risk Control and Reliability: UAT (Unsupervised Adaptive Thresholding) designs a reliability function for early-exit DNNs to assess the quality of intermediate layer outputs, and employs a multi-armed bandit (MAB) algorithm to dynamically learn optimal exit thresholds at inference time, achieving 1.7–2.1× speedup with less than 2% performance degradation while remaining robust to distribution shift.

🎵 Audio & Speech (47)¶

VITA-1.5: Towards GPT-4o Level Real-Time Vision and Speech Interaction: VITA-1.5 proposes a carefully designed three-stage progressive training strategy that incrementally integrates visual and speech capabilities into an LLM, achieving end-to-end vision-speech real-time interaction without relying on standalone ASR/TTS modules, while attaining state-of-the-art performance among open-source models on image, video, and speech benchmarks.
LeVo: High-Quality Song Generation with Multi-Preference Alignment: This paper proposes LeVo, a song generation framework that employs a language model to jointly model mixed tokens and dual-track tokens, thereby reconciling vocal-accompaniment harmony with audio quality. It further introduces a DPO-based multi-preference alignment method to enhance musicality and instruction-following capability.
MEGADance: Mixture-of-Experts Architecture for Genre-Aware 3D Dance Generation: This paper proposes MEGADance, the first music-driven 3D dance generation method based on a Mixture-of-Experts (MoE) architecture. It decouples choreographic consistency into "dance universality" (Universal Expert) and "style specificity" (Specialized Expert), combined with FSQ quantization and a Mamba-Transformer hybrid backbone, achieving state-of-the-art dance quality and strong style controllability.
Seeing Sound, Hearing Sight: Uncovering Modality Bias and Conflict of AI Models in Sound Localization: This work systematically reveals that AI SSL models suffer from severe visual bias—degrading to near-random performance under audio-visual conflict—and proposes EchoPin, a neuroscience-inspired model (HRTF filtering + cochleagram + stereo audio) that substantially outperforms prior methods on AudioCOCO and exhibits a human-like horizontal-over-vertical localization accuracy asymmetry.
AudSemThinker: Enhancing Audio-Language Models through Reasoning over Semantics of Sound: AudSemThinker introduces a structured semantic reasoning framework for audio-language models by defining 9 categories of sound semantic descriptors (who/what/how/when/where, etc.). Built on Qwen2.5-Omni-7B and trained via SFT + GRPO (with verifiable rewards and length constraints), the model produces three-stage outputs in the format \<think>\<semantic_elements>\<answer>, achieving 66.70% on the MMAU benchmark—surpassing Audio-Reasoner (61.71%) and Qwen2.5-Omni (65.60%).
A TRIANGLE Enables Multimodal Alignment Beyond Cosine Similarity: TRIANGLE proposes using the area of the triangle formed by three modal embedding vectors in high-dimensional space as a similarity measure, replacing traditional pairwise cosine similarity to achieve joint alignment of video, audio, and text modalities. The method surpasses state-of-the-art by up to 9 Recall@1 points on video-text retrieval and related tasks.
E-BATS: Efficient Backpropagation-Free Test-Time Adaptation for Speech Foundation Models: This paper proposes E-BATS, the first backpropagation-free test-time adaptation framework for speech foundation models. Through lightweight prompt adaptation, multi-scale loss functions, and a test-time EMA mechanism, E-BATS achieves 2.0×–6.4× GPU memory savings while maintaining competitive accuracy.
SHAP Meets Tensor Networks: Provably Tractable Explanations with Parallelism: This paper presents the first provably exact SHAP computation framework for Tensor Networks (TNs), proves that SHAP under the Tensor Train (TT) structure is parallelizable in polylogarithmic time (NC² complexity), and reveals via reductions that width rather than depth is the fundamental bottleneck for SHAP computation in binarized neural networks.
Brain-tuning Improves Generalizability and Efficiency of Brain Alignment in Speech Models: This paper proposes Multi-brain-tuning, a method that jointly fine-tunes pretrained speech models on fMRI data from multiple participants, reducing the data required for brain alignment by 5×, improving alignment by up to 50%, and generalizing to unseen participants and datasets.
Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation: This paper proposes RecBench, a comprehensive evaluation framework that systematically compares 17 LLMs against 10 conventional DLRMs across 5 domain-specific datasets. Results show that LLM-based recommenders achieve up to 5% AUC improvement on CTR tasks and up to 170% NDCG@10 improvement on sequential recommendation, yet incur 10–1000× slower inference. Conventional DLRMs augmented with LLM semantic embeddings (LLM-for-RS) attain approximately 95% of LLM performance at 20× higher throughput, making this paradigm the most industrially viable solution at present.

Browse all 47 Audio & Speech papers →

🔎 AIGC Detection (9)¶

Classical Planning with LLM-Generated Heuristics: Challenging the State of the Art with Python Code: This paper proposes having LLMs generate Python code for domain-dependent heuristic functions (rather than directly generating plans). Candidate heuristics are obtained via $n$ samples and the best is selected on a training set, then injected into the Python planner Pyperplan for use with GBFS. The approach surpasses all C++ Fast Downward traditional heuristics on 8 IPC 2023 benchmark domains using pure Python, matches the SOTA learned planner $h^{\mathrm{WLF}}_{\mathrm{GPR}}$, and guarantees 100% correctness for all plans found.
Can LLMs Write Faithfully? An Agent-Based Evaluation of LLM-generated Islamic Content: This paper proposes a dual-agent (quantitative + qualitative) evaluation framework that systematically assesses the faithfulness of GPT-4o, Ansari AI, and Fanar on Islamic content generation tasks across three dimensions—theological accuracy, citation integrity, and stylistic appropriateness—finding that even the best-performing model exhibits significant deficiencies in citation reliability.
"Jutters": Through the metaphor of the Dutch tradition of jutters (beachcombers), this work constructs an immersive installation art piece that integrates real beach debris with AI-generated images and videos, guiding visitors to adopt a beachcomber's mindset in reflecting on how to engage with AI-generated content.
ASCIIBench: Evaluating Language-Model-Based Understanding of Visually-Oriented Text: This paper introduces ASCIIBench, the first publicly available benchmark for ASCII art understanding and generation (5,315 images, 752 categories). Systematic evaluation reveals that the visual modality substantially outperforms the text modality, multimodal fusion yields no benefit, and CLIP exhibits a fundamental bottleneck in representing ASCII structure—only categories with high intra-class consistency can be effectively distinguished.
CLAWS: Creativity Detection for LLM-Generated Solutions Using Attention Window of Sections: This paper proposes CLAWS, a method that analyzes the attention weight distribution of LLMs across different prompt sections during mathematical solution generation to classify outputs as "creative," "typical," or "hallucinated," without requiring human evaluation.
DuoLens: A Framework for Robust Detection of Machine-Generated Multilingual Text and Code: DuoLens is proposed — an AI-generated content detection framework based on dual-encoder fusion of CodeBERT and CodeBERTa — achieving AUROC of 0.97–0.99 on multilingual text (8 languages) and source code (7 programming languages) at significantly reduced computational cost (8–12× lower latency, 3–5× lower VRAM), substantially outperforming large models such as GPT-4o.
QiMeng-NeuComBack: Self-Evolving Translation from IR to Assembly Code: This paper introduces the NeuComBack benchmark for evaluating neural compilation on IR-to-assembly translation tasks, and proposes a self-evolving prompt optimization method that iteratively improves compilation prompts by learning from LLM self-debugging trajectories. The approach raises correctness from 44% to 64%, with 87.5% of correctly generated programs outperforming clang-O3.
Reasoning Compiler: LLM-Guided Optimizations for Efficient Model Serving: This paper proposes Reasoning Compiler, which models compiler optimization as a sequential decision-making process, employing an LLM as a context-aware proposal engine combined with MCTS to balance exploration and exploitation. The approach achieves an average 5.0× speedup across 5 representative benchmarks and 5 hardware platforms, with 10.8× better sampling efficiency than TVM's evolutionary search.
Synthesizing Performance Constraints for Evaluating and Improving Code Efficiency: This paper proposes Wedge, a framework that uses LLMs to synthesize performance-characterizing constraints to guide constraint-aware fuzzing, generating stress-test inputs that expose code performance bottlenecks. It further constructs the PerfForge benchmark, enabling LLM-based code optimizers (e.g., Effi-Learner) to achieve up to 24% additional reduction in CPU instructions.

🧊 3D Reconstruction & Rendering (40)¶

4DGT: Learning a 4D Gaussian Transformer Using Real-World Monocular Videos: This paper proposes 4DGT — a 4D Gaussian-based Transformer model trained entirely on real-world monocular posed videos that reconstructs dynamic scenes in seconds via feed-forward inference, significantly outperforming comparable feed-forward networks while achieving accuracy on par with optimization-based methods.
Anti-Aliased 2D Gaussian Splatting: This paper proposes AA-2DGS, which addresses severe aliasing artifacts in 2D Gaussian Splatting under varying sampling rates through two complementary mechanisms: a world-space flat smoothing kernel and an object-space Mip filter. The method significantly improves multi-scale rendering quality while preserving the geometric accuracy advantages of 2DGS.
AtlasGS: Atlanta-world Guided Surface Reconstruction with Implicit Structured Gaussians: AtlasGS is proposed to achieve smooth, high-frequency-detail-preserving surface reconstruction in indoor and urban scenes by incorporating the Atlanta-world structural prior into an implicit-structured Gaussian representation, comprehensively outperforming existing implicit and explicit methods.
D$^2$USt3R: Enhancing 3D Reconstruction for Dynamic Scenes: This paper proposes the Static-Dynamic Aligned Pointmap (SDAP) representation, which unifies 3D alignment of static and dynamic regions into a single framework, enabling DUSt3R-based methods to achieve accurate dense 3D reconstruction and correspondence estimation in dynamic scenes.
DC4GS: Directional Consistency-Driven Adaptive Density Control for 3D Gaussian Splatting: This paper proposes DC4GS, an adaptive density control method based on Directional Consistency (DC), which improves primitive splitting decisions and split position selection in 3DGS by exploiting the angular coherence of positional gradients. DC4GS reduces the number of primitives by up to 30% while improving reconstruction quality.
Dynamic Gaussian Splatting from Defocused and Motion-blurred Monocular Videos: A unified framework is proposed that jointly models defocus blur and motion blur via learnable blur kernel convolution, combined with a dynamic Gaussian densification strategy and unseen-view constraints, enabling high-quality novel view synthesis of dynamic scenes from blurry monocular videos using 3DGS.
EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting: EF-3DGS is the first work to introduce event cameras into free-trajectory scene reconstruction. It employs an Event Generation Model (EGM) to reconstruct latent inter-frame images for continuous supervision, Contrast Maximization (CMax) combined with a Linear Event Generation Model (LEGM) to extract motion information for pose calibration, and a photometric BA + Fixed-GS strategy to resolve color inconsistency. The method achieves a 3 dB PSNR improvement and a 40% reduction in ATE in high-speed scenarios.
Every Camera Effect, Every Time, All at Once: 4D Gaussian Ray Tracing for Physics-based Camera Effect Data Generation: This paper proposes 4D Gaussian Ray Tracing (4D-GRT), which integrates 4D Gaussian Splatting with physics-based ray tracing. After reconstructing dynamic scenes from multi-view videos, the method renders physically accurate video data with controllable camera effects including fisheye distortion, depth of field blur, and rolling shutter artifacts.
Flux4D: Flow-based Unsupervised 4D Reconstruction: Flux4D is proposed as an unsupervised and generalizable 4D dynamic driving scene reconstruction framework. It employs a feed-forward network to directly predict 3D Gaussians and their motion velocities, achieving large-scale scene reconstruction using only photometric loss and a static-preference regularization. The method surpasses all unsupervised approaches on PandaSet and Waymo while approaching the performance of supervised methods.
GeoSVR: Taming Sparse Voxels for Geometrically Accurate Surface Reconstruction: This paper proposes GeoSVR, an explicit surface reconstruction framework based on sparse voxels. By introducing voxel-uncertainty depth constraints and sparse voxel surface regularization, GeoSVR comprehensively outperforms existing 3DGS- and SDF-based methods in geometric accuracy, detail preservation, and reconstruction completeness.

Browse all 40 3D Reconstruction & Rendering papers →

🧊 3D Generation & Editing (12)¶

ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction: This paper proposes to formulate 3D mesh generation as a coarse-to-fine, next-level-of-detail prediction process. By reversing a generalized mesh simplification algorithm (GSlim), a progressive refinement sequence is obtained, which is then learned autoregressively via a Transformer. Generation begins from a single point and incrementally adds geometric and topological detail to produce a complete mesh.
Cue3D: Quantifying the Role of Image Cues in Single-Image 3D Generation: Cue3D is the first model-agnostic framework for quantifying the importance of image cues in single-image 3D generation. By systematically perturbing six visual cues—illumination, texture, silhouette, perspective, edges, and local continuity—across seven methods spanning three paradigms (regression-based, multi-view, and native 3D generation), it reveals key insights: shape meaningfulness rather than texture governs generalization ability, illumination matters more than texture, and models are overly dependent on input silhouettes.
From Programs to Poses: Factored Real-World Scene Generation via Learned Program Libraries: This paper proposes FactoredScenes, which decomposes real-world 3D scene generation into a five-step factorization pipeline — learning a layout program library from synthetic data, generating scene programs via LLM, executing programs to obtain axis-aligned layouts, program-conditioned hierarchical pose prediction, and object retrieval and placement. The method achieves 38.3% FID improvement and 80.4% KID improvement on bedrooms, with human evaluators able to distinguish generated scenes from real ScanNet scenes only 67% of the time.
GOATex: Geometry & Occlusion-Aware Texturing: GOATex proposes the first occlusion-aware 3D mesh texturing framework. It decomposes meshes into visibility layers ordered from outermost to innermost via a ray-casting-based hit-level mechanism, applies a two-stage visibility control strategy combining normal flipping and residual face clustering, and performs visibility-weighted blending in UV space—achieving high-quality texture generation for both exterior surfaces and occluded interior surfaces.
Mesh-RFT: Enhancing Mesh Generation via Fine-Grained Reinforcement Fine-Tuning: This paper proposes Mesh-RFT, a framework that achieves face-level fine-grained mesh quality optimization through a topology-aware scoring system and Masked Direct Preference Optimization (M-DPO), significantly improving the geometric integrity and topological regularity of generated meshes.
Orientation Matters: Making 3D Generative Models Orientation-Aligned: This paper introduces the task of orientation-aligned 3D object generation, constructs the Objaverse-OA dataset comprising 14,832 orientation-aligned 3D models across 1,008 categories, fine-tunes two mainstream 3D generation frameworks (Trellis and Wonder3D) to achieve orientation-aligned object generation, and demonstrates two downstream applications: zero-shot orientation estimation and arrow-guided rotation manipulation.
PhysX-3D: Physical-Grounded 3D Asset Generation: PhysX proposes the first end-to-end physical-property-driven 3D asset generation paradigm, comprising PhysXNet (the first 3D dataset with systematic annotations across five physical dimensions—absolute scale, material, functional affordance, kinematics, and functional description—covering 26K+ objects) and PhysXGen (a dual-branch feed-forward generation framework that injects physical knowledge into a pretrained 3D structural latent space).
PointMAC: Meta-Learned Adaptation for Robust Test-Time Point Cloud Completion: PointMAC is the first framework to introduce meta-auxiliary learning and test-time adaptation (TTA) into point cloud completion. It leverages Bi-Aux Units (random masked reconstruction + denoising) as self-supervised signals, employs MAML to align auxiliary objectives with the primary task, and at inference updates only the shared encoder for sample-level refinement, achieving state-of-the-art performance on synthetic, simulated, and real-world data.
Scaffold Diffusion: Sparse Multi-Category Voxel Structure Generation with Discrete Diffusion: This paper proposes Scaffold Diffusion, which treats sparse multi-category 3D voxels as token sequences and employs a Masked Diffusion Language Model (MDLM) with 3D sinusoidal positional encodings to generate spatially coherent multi-category voxel structures conditioned on occupancy maps. The method substantially outperforms autoregressive and conventional discrete diffusion baselines on an extremely sparse (>98% background) Minecraft house dataset.
SceneWeaver: All-in-One 3D Scene Synthesis with an Extensible and Self-Reflective Agent: This paper proposes SceneWeaver, the first reflective agent framework for 3D scene synthesis, which unifies multiple scene generation paradigms through a standardized and extensible tool interface. By employing a reason-act-reflect closed-loop for iterative refinement, it comprehensively outperforms existing methods in physical plausibility, visual realism, and semantic alignment.

Browse all 12 3D Generation & Editing papers →

🧊 3D Representation & Scene Understanding (13)¶

EA3D: Online Open-World 3D Object Extraction from Streaming Videos: This paper proposes EA3D (ExtractAnything3D), an online open-world 3D object extraction framework that performs simultaneous geometric reconstruction and comprehensive scene understanding from streaming videos via knowledge-integrated feature maps, online visual odometry, and recurrent joint optimization.
From Objects to Anywhere: A Holistic Benchmark for Multi-level Visual Grounding in 3D Scenes: This paper proposes Anywhere3D-Bench, the first 3D visual grounding benchmark spanning four levels—area, space, object, and part—revealing that even the strongest models (Gemini-2.5-Pro and o3) achieve only ~30% accuracy on space-level tasks and ~40% on part-level tasks, far below the human performance of 95%.
HouseLayout3D: A Benchmark and Training-Free Baseline for 3D Layout Estimation in the Wild: This paper introduces HouseLayout3D—the first real-world 3D layout estimation benchmark targeting large-scale multi-floor buildings—and MultiFloor3D, a training-free baseline that combines modern 3D reconstruction and segmentation models to surpass existing deep learning methods on multi-floor building layout estimation.
LangSplatV2: High-dimensional 3D Language Gaussian Splatting with 450+ FPS: By treating each 3D Gaussian as a sparse code over a global dictionary, LangSplatV2 replaces the heavyweight decoder with a sparse coefficient field, achieving 476.2 FPS high-dimensional feature splatting and 384.6 FPS 3D open-vocabulary querying — a 47× speedup over LangSplat.
Learning Generalizable Shape Completion with SIM(3) Equivariance: This paper proposes SIMECO, the first SIM(3)-equivariant shape completion network. Through a three-stage modular design — feature canonicalization → similarity-invariant geometric reasoning → transformation recovery — SIMECO outperforms all augmentation-based and equivariant baselines under an unbiased evaluation protocol, achieving a 17% MMD reduction on KITTI and a 14% CD-$\ell_1$ reduction on OmniObject3D. Notably, SIMECO under the stricter protocol still surpasses competing methods evaluated under their own biased settings.
Motion4D: Learning 3D-Consistent Motion and Semantics for 4D Scene Understanding: Motion4D proposes a unified 4D Gaussian splatting framework that incorporates priors from 2D foundation models (semantic masks, point tracking, depth) into 3D representations via an iterative refinement strategy, achieving spatiotemporally consistent motion and semantic modeling. The method significantly outperforms existing approaches on video object segmentation, point tracking, and novel view synthesis tasks.
Novel Class Discovery for Point Cloud Segmentation via Joint Learning of Causal Representation and Reasoning: This paper is the first to introduce causal learning into 3D point cloud novel class discovery (3D-NCD). By leveraging a Structural Causal Model (SCM) to analyze confounders in base classes and causal relationships between base and novel classes, it proposes Causal Representation Prototype learning (CRP, which eliminates confounders via an adversarial network) and graph-based causal reasoning (GCN-based pseudo-label generation), achieving state-of-the-art results on SemanticKITTI and SemanticPOSS.
Object-Centric Representation Learning for Enhanced 3D Semantic Scene Graph Prediction: Through empirical analysis, this paper identifies object feature discriminability as the critical bottleneck in 3D scene graph predicate prediction (object misclassification accounts for 92%+ of predicate errors). It proposes an independently contrastively pre-trained object encoder (3D-2D-Text tri-modal alignment), a geometry-regularized relation encoder, and a bidirectional edge-gated GNN, achieving new SOTA on 3DSSG with Object R@1 59.53% and Predicate R@50 91.40%.
Online Segment Any 3D Thing as Instance Tracking: AutoSeg3D reformulates online 3D instance segmentation as an instance tracking problem, leveraging long-term memory for cross-frame instance association, short-term memory for instance update, and spatial consistency learning to mitigate VFM over-segmentation. The method surpasses ESAM by 2.8 AP on ScanNet200 while maintaining real-time performance.
OpenLex3D: A Tiered Evaluation Benchmark for Open-Vocabulary 3D Scene Representations: This paper proposes OpenLex3D, a tiered evaluation benchmark for open-vocabulary 3D scene representations. Built upon Replica, ScanNet++, and HM3D, it provides language annotations 13× richer than the original labels, supporting evaluation on two tasks: open-set 3D semantic segmentation and object retrieval.

Browse all 13 3D Representation & Scene Understanding papers →

🧊 Depth, Normals & Stereo (5)¶

3D Visual Illusion Depth Estimation: This paper reveals that 3D visual illusions (e.g., wall paintings, screen replays, mirror reflections) severely mislead existing state-of-the-art monocular and stereo depth estimation methods. The authors construct a large-scale dataset comprising approximately 3k scenes and 200k images, and propose a VLM-driven monocular-stereo adaptive fusion framework that achieves state-of-the-art performance across diverse illusion scenarios.
DualFocus: Depth from Focus with Spatio-Focal Dual Variational Constraints: This paper proposes DualFocus, which achieves robust and accurate depth estimation from focal stacks via two complementary constraints: a spatial variational constraint (exploiting focus-dependent gradient patterns to distinguish depth edges from texture artifacts) and a focal variational constraint (enforcing a unimodal and monotonic focus probability distribution along the focal axis).
Jasmine: Harnessing Diffusion Prior for Self-Supervised Depth Estimation: This paper is the first to incorporate the visual prior of Stable Diffusion into a self-supervised monocular depth estimation (SSMDE) framework. It proposes the Mix-Batch Image Reconstruction (MIR) proxy task to shield the SD prior from corruption by reprojection noise, and introduces the Scale-Shift GRU (SSG) to bridge the gap between SD's scale-shift-invariant (SSI) and self-supervised scale-invariant (SI) depth distributions. Jasmine achieves AbsRel = 0.090 on KITTI, establishing a new state of the art among all SSMDE methods, while comprehensively outperforming supervised SD methods such as Marigold, E2E FT, and Lotus in zero-shot generalization.
MVSMamba: Multi-View Stereo with State Space Model: This paper proposes MVSMamba, the first Mamba-based Multi-View Stereo (MVS) network, which achieves efficient intra-view and inter-view global omnidirectional feature aggregation via a reference-centered dynamic scanning strategy, attaining state-of-the-art performance on DTU and Tanks-and-Temples with superior efficiency.
Pixel-Perfect Depth with Semantics-Prompted Diffusion Transformers: This paper proposes Pixel-Perfect Depth, a monocular depth estimation model that performs diffusion generation directly in pixel space (rather than latent space). Through a Semantics-Prompted DiT (SP-DiT) that incorporates high-level semantic representations from visual foundation models and a cascaded DiT design, the model generates flying-pixel-free depth maps, surpassing all published generative models on five benchmarks.

🧊 Geometric Matching & Localization (6)¶

E-MoFlow: Learning Egomotion and Optical Flow from Event Data via Implicit Regularization: This paper proposes E-MoFlow, which models optical flow as an implicit neural representation and egomotion as a continuous spline, jointly optimizing both via differential geometric constraints under an unsupervised paradigm to achieve 6-DoF egomotion and dense optical flow estimation from event data.
Locality-Sensitive Hashing-Based Efficient Point Transformer for Charged Particle Reconstruction: By combining LSH with Point Transformer, the paper proposes HEPTv2 for end-to-end particle track reconstruction, eliminating the DBScan clustering post-processing bottleneck and achieving a 28.9× speedup while maintaining competitive tracking efficiency.
Rectified Point Flow: Generic Point Cloud Pose Estimation: This paper proposes Rectified Point Flow, a unified generative framework that reformulates pairwise point cloud registration and multi-part shape assembly as a conditional generation problem, estimating part poses by learning a continuous point-wise velocity field.
RGB-Only Supervised Camera Parameter Optimization in Dynamic Scenes: ROS-Cam proposes a camera parameter (focal length + pose) optimization method for dynamic scenes supervised solely by a single RGB video. It achieves state-of-the-art accuracy and fastest runtime across 5 datasets via three key contributions: a patch-wise tracking filter for sparse, robust correspondences; a Cauchy distribution-based outlier-aware joint optimization that adaptively down-weights moving objects; and a two-stage optimization strategy grounded in Softplus/convex minimax analysis.
SingRef6D: Monocular Novel Object Pose Estimation with a Single RGB Reference: SingRef6D is a lightweight 6D pose estimation pipeline requiring only a single RGB reference image. It fine-tunes Depth-Anything v2 via a token-scaler mechanism for robust depth prediction and introduces depth-aware matching to enhance LoFTR's spatial reasoning, substantially outperforming existing methods on transparent and reflective objects.
TAPIP3D: Tracking Any Point in Persistent 3D Geometry: This paper proposes TAPIP3D, which represents video as a camera-stabilized spatiotemporal 3D feature point cloud and iteratively refines multi-frame point trajectories in persistent 3D geometric space via a 3D Neighborhood-to-Neighborhood (N2N) attention mechanism, substantially outperforming existing 3D point tracking methods.

🎯 Object Detection (27)¶

Video-RAG: Visually-aligned Retrieval-Augmented Long Video Comprehension: This paper proposes Video-RAG, a training-free, plug-and-play RAG pipeline that extracts visually-aligned auxiliary texts (OCR, ASR, object detection) from video, retrieves relevant content, and feeds it into LVLMs. With an overhead of only ~2K tokens, it improves average Video-MME performance by 2.8% across seven open-source LVLMs, and the 72B model surpasses GPT-4o.
Normal-Abnormal Guided Generalist Anomaly Detection: NAGL is the first framework to incorporate mixed normal-and-abnormal reference samples into Generalist Anomaly Detection (GAD). Through two attention modules—Residual Mining (RM) and Anomaly Feature Learning (AFL)—it learns transferable anomaly patterns in residual space, substantially outperforming normal-reference-only methods in cross-domain scenarios with as few as 1 anomaly reference sample.
Scalable, Explainable and Provably Robust Anomaly Detection with One-Step Flow Matching: This paper proposes TCCM (Time-Conditioned Contraction Matching), a flow matching-inspired semi-supervised anomaly detection method for tabular data. By learning a time-conditioned velocity field that contracts normal data toward the origin, TCCM computes anomaly scores in a single forward pass, achieving top AUROC and AUPRC rankings across 47 ADBench datasets while running 1573× faster than DTE.
DETree: DEtecting Human-AI Collaborative Texts via Tree-Structured Hierarchical Representation Learning: This paper proposes DETree, a framework that constructs a Hierarchical Affinity Tree (HAT) to model the hierarchical relationships among diverse human-AI collaborative text generation processes, and designs a Tree-Structured Contrastive Loss (TSCL) to align the representation space. DETree achieves significant advantages in mixed-text detection and OOD generalization scenarios.
Ascent Fails to Forget: Starting from the statistical dependence between the forget set and the retain set, this paper theoretically and empirically demonstrates that the widely adopted gradient ascent / Descent-Ascent (DA) family of machine unlearning methods fails systematically in the presence of data correlations. In logistic regression, the DA solution is provably farther from the oracle than the original model, and in non-convex settings DA traps the model in inferior local minima.
ReCon-GS: Continuum-Preserved Gaussian Streaming for Fast and Compact Reconstruction: This paper proposes ReCon-GS, which achieves incremental 3D reconstruction via continuum-preserved Gaussian streaming, substantially reducing storage requirements and training time while maintaining rendering quality, and supporting real-time reconstruction of large-scale scenes.
DCAD-2000: A Multilingual Dataset across 2000+ Languages with Data Cleaning as Anomaly Detection: This work constructs DCAD-2000, a multilingual dataset covering 2,282 languages and 46.72 TB of text, and proposes a language-agnostic data cleaning framework that reformulates cleaning as anomaly detection. The framework extracts 8-dimensional statistical features per document and applies Isolation Forest for dynamic noise filtering. Effectiveness is validated on multiple multilingual benchmarks, with particularly notable gains on low-resource languages.
Rethinking Evaluation of Infrared Small Target Detection: This paper systematically identifies three critical limitations in existing evaluation protocols for infrared small target detection (IRSTD), and proposes a hierarchical analysis framework comprising the hybrid-level metric hIoU, a systematic error analysis methodology, and a cross-dataset evaluation setting.
Semi-supervised Graph Anomaly Detection via Robust Homophily Learning: This paper proposes RHO (Robust Homophily Learning), which addresses the homophily diversity of normal nodes in semi-supervised graph anomaly detection via an adaptive frequency response filter (AdaFreq) and a Graph Normality Alignment (GNA) module, outperforming existing methods on 8 real-world datasets.
Test-Time Adaptive Object Detection with Foundation Model: This paper proposes TTAOD, a source-free open-vocabulary test-time adaptive object detection framework that combines multimodal Prompt Tuning, Mean-Teacher, an Instance Dynamic Memory (IDM) module, and memory augmentation/hallucination strategies. It achieves 56.2% AP50 on Pascal-C (+11.0 vs. SOTA) and demonstrates consistent gains across 13 cross-domain datasets.

Browse all 27 Object Detection papers →

✂️ Segmentation (45)¶

UniPixel: Unified Object Referring and Segmentation for Pixel-Level Visual Reasoning: UniPixel proposes the first end-to-end large multimodal model that unifies object referring and segmentation, leveraging a novel Object Memory Bank design to transform sparse visual prompts into dense object mask features and inject them into the reasoning process. The model achieves state-of-the-art performance on 10 benchmarks and introduces PixelQA, a new task requiring simultaneous referring, segmentation, and question answering.
SAM-R1: Leveraging SAM for Reward Feedback in Multimodal Segmentation via Reinforcement Learning: SAM-R1 proposes an end-to-end reasoning segmentation framework that, for the first time, incorporates SAM as a reward provider within the reinforcement learning training loop. Combined with a tiered IoU accuracy reward, asymmetric clipping, and token-level loss normalization in an improved GRPO algorithm, the method achieves a gIoU of 60.2% on the ReasonSeg zero-shot benchmark—surpassing Seg-Zero and other approaches—using only 3K training samples.
PartNeXt: A Next-Generation Dataset for Fine-Grained and Hierarchical 3D Part Understanding: This paper presents PartNeXt, a fine-grained hierarchical part annotation dataset comprising 23,519 high-quality textured 3D models across 50 categories. Two benchmarks are established—category-agnostic part segmentation and 3D part question answering—revealing significant deficiencies of current methods in fine-grained part understanding.
Vanish into Thin Air: Cross-prompt Universal Adversarial Attacks for SAM2: This paper proposes UAP-SAM2—the first cross-prompt universal adversarial attack against SAM2—which employs a dual semantic shift framework (intra-frame semantic confusion + inter-frame semantic inconsistency) to generate a single universal perturbation that causes segmentation targets to "vanish" across different videos, frames, and prompts.
RoMA: Scaling up Mamba-based Foundation Models for Remote Sensing: This paper proposes RoMA — the first self-supervised autoregressive pre-training framework based on the Mamba architecture for remote sensing. By introducing an adaptive rotation encoding strategy and a multi-scale token prediction mechanism, RoMA addresses the challenges of orientation diversity and extreme scale variation inherent in remote sensing imagery, while empirically validating that Mamba follows data and parameter scaling laws in the remote sensing domain.
Attention (as Discrete-Time Markov) Chains: This work reinterprets the softmax-normalized attention matrix as the transition probability matrix of a Discrete-Time Markov Chain (DTMC), and proposes Multi-Bounce Attention and TokenRank (stationary distribution, analogous to PageRank) to capture indirect attention paths and global token importance. The approach achieves 94.29% mAP on ImageNet segmentation and enhances image generation quality in Self-Attention Guidance.
COS3D: Collaborative Open-Vocabulary 3D Segmentation: This paper proposes COS3D — a collaborative prompt-segmentation framework that constructs a collaborative field comprising an instance field and a language field. During training, the language field is built via instance-to-language feature mapping; during inference, language-to-instance adaptive prompt refinement generates precise segmentation results. COS3D substantially outperforms existing methods on two mainstream benchmarks.
HopaDIFF: Holistic-Partial Aware Fourier Conditioned Diffusion for Referring Human Action Segmentation in Multi-Person Scenarios: This paper introduces the Referring Human Action Segmentation (RHAS) task—localizing a specific individual in multi-person videos via textual descriptions and performing frame-level action segmentation. The authors construct the RHAS133 dataset comprising 133 movies, 137 action categories, and 33 hours of video, and propose HopaDIFF, a holistic-partial aware Fourier-conditioned diffusion framework that substantially outperforms existing baselines across multiple evaluation settings.
InstructSAM: A Training-Free Framework for Instruction-Oriented Remote Sensing Object Recognition: This paper introduces a new task — Instruction-oriented Counting, Detection, and Segmentation (InstructCDS) — along with the EarthInstruct remote sensing benchmark covering three settings (open-vocabulary, open-ended, and open-subcategory). It proposes InstructSAM, a training-free framework that uses an LVLM to parse instructions and predict counts, SAM2 to generate mask proposals, and CLIP to compute similarities. A Binary Integer Programming (BIP) formulation then performs optimal mask-label assignment under counting constraints, achieving near-constant inference time while outperforming task-specific baselines.
Alligat0R: Pre-Training through Covisibility Segmentation for Relative Camera Pose Regression: This paper replaces CroCo's cross-view completion with covisibility segmentation as a stereo vision pre-training task, predicting per-pixel labels of "co-visible / occluded / out-of-view" for each pixel. The approach significantly outperforms CroCo in low-overlap scenarios and achieves a first-place overall success rate of 60.3% on the RUBIK benchmark.

Browse all 45 Segmentation papers →

🖼️ Image Restoration (29)¶

DynaGuide: Steering Diffusion Policies with Active Dynamic Guidance: This paper proposes DynaGuide, which applies classifier guidance to a frozen pretrained diffusion policy at inference time via an external latent dynamics model, steering the robot toward arbitrary positive/negative goals without modifying policy weights. It achieves an average success rate of 70% on CALVIN simulation and 80% on a real robot.
Spiking Meets Attention: Efficient Remote Sensing Image Super-Resolution with Attention Spiking Neural Networks: This paper proposes SpikeSR, the first attention-based spiking neural network (SNN) framework for remote sensing image super-resolution. By incorporating Spiking Attention Blocks (SAB) that combine Hybrid Dimensional Attention (HDA) and Deformable Similarity Attention (DSA), SpikeSR achieves state-of-the-art performance on AID/DOTA/DIOR while maintaining high computational efficiency.
Encoder-Decoder Diffusion Language Models for Efficient Training and Inference: This paper proposes E2D2, an encoder-decoder architecture for discrete diffusion language models that performs iterative denoising via a lightweight decoder while periodically updating representations through a large encoder, achieving faster inference (~3× vs. MDLM) and more efficient block diffusion training (halving FLOPs).
Audio Super-Resolution with Latent Bridge Models: This paper proposes AudioLBM, which compresses audio waveforms into a continuous latent space and employs a bridge model to realize a latent-to-latent generation process from low-resolution to high-resolution. Combined with frequency-aware training for broader data utilization and a cascaded design to surpass the 48kHz ceiling, AudioLBM comprehensively outperforms methods such as AudioSR across speech, sound effects, and music, while achieving any-to-192kHz audio super-resolution for the first time.
MS-BART: Unified Modeling of Mass Spectra and Molecules for Structure Elucidation: This paper proposes MS-BART, which maps molecular fingerprints and molecular structures (SELFIES) into a shared token space via a unified vocabulary, performs multi-task pretraining on 4 million fingerprint–molecule pairs, and subsequently applies experimental spectra fine-tuning and chemical feedback alignment to enable efficient generation of molecular structures from mass spectra.
Rethinking Circuit Completeness in Language Models: AND, OR, and ADDER Gates: This paper systematically introduces AND, OR, and ADDER gates to decompose language model circuits, reveals that circuit incompleteness primarily stems from the omission of OR gates, and proposes a framework combining noising and denoising interventions to fully recover all three gate types while guaranteeing both faithfulness and completeness.
RGB-to-Polarization Estimation: A New Task and Benchmark Study: This paper formally defines the novel task of estimating polarization components (S₁/S₂/S₃) from standard RGB images, establishes the first systematic benchmark encompassing both restoration-based and generative methods, and finds that pretrained MAE achieves the best overall pixel-level accuracy (PSNR 24.74). Restoration-based methods consistently outperform diffusion-based generative methods, with pretrained weight transfer identified as a critical advantage.
Rethinking Nighttime Image Deraining via Learnable Color Space Transformation: Motivated by the statistical finding that nighttime rain exhibits far greater contrast in the Y channel (luminance) of YCbCr than in RGB, this work proposes a learnable Color Space Converter (CSC) that performs deraining in the Y channel, an Implicit Illumination Guidance (IIG) module that encodes non-uniform nighttime illumination, and a photorealistic dataset HQ-NightRain constructed via illumination-aware synthesis. The three components jointly yield substantial improvements in nighttime deraining performance.
MoDEM: A Morton-Order Degradation Estimation Mechanism for Adverse Weather Image Restoration: This paper proposes the MODEM framework, which combines Morton-encoded spatial scanning with selective state space models (SSMs) to capture spatially heterogeneous weather degradation patterns. Equipped with a dual degradation estimation module that provides both global and local priors, MODEM achieves state-of-the-art unified adaptive restoration across multiple adverse weather degradation types.
MAP Estimation with Denoisers: Convergence Rates and Guarantees: This paper proves that a simple iterative averaging algorithm based on MMSE denoisers—closely related to practical methods such as Cold Diffusion—provably converges to the proximal operator of the negative log-prior under log-concave prior assumptions, achieving a convergence rate of $\tilde{O}(1/k)$. The work provides rigorous theoretical foundations for a class of denoising methods that have demonstrated empirical success but lacked theoretical guarantees, and embeds the approach within a proximal gradient descent framework for MAP estimation.

Browse all 29 Image Restoration papers →

🛰️ Remote Sensing (12)¶

RSCC: A Large-Scale Remote Sensing Change Caption Dataset for Disaster Events: This work introduces RSCC — the first large-scale disaster-aware remote sensing change captioning dataset comprising 62,351 pre/post-disaster image pairs with detailed change descriptions, covering 31 global disaster events including earthquakes, floods, and wildfires. High-quality annotations are generated using the QvQ-Max visual reasoning model, and a comprehensive benchmark evaluation framework is established.
OrthoLoC: UAV 6-DoF Localization and Calibration Using Orthographic Geodata: OrthoLoC establishes the first large-scale UAV 6-DoF localization benchmark dataset based on orthographic geodata (DOP+DSM), comprising 16,425 real UAV images across 47 regions in Germany and the United States. It further introduces AdHoP (Adaptive Homography Preprocessing), a matching enhancement technique that improves matching performance by 95% and reduces translation error by 63% without modifying the underlying feature matcher.
ChA-MAEViT: Unifying Channel-Aware Masked Autoencoders and Multi-Channel Vision Transformers for Improved Cross-Channel Learning: This paper proposes ChA-MAEViT, which enhances cross-channel feature learning for multi-channel images (MCI) through four key components: dynamic channel-patch joint masking, memory tokens, hybrid token fusion, and a channel-aware decoder. The method outperforms the state of the art by an average of 3.0–21.5% across three satellite and microscopy datasets.
Scaling Image Geo-Localization to Continent Level: A hybrid approach combining classification-learned prototypes with aerial image embeddings achieves 68%+ recall@1 within 200 m and 59.2% within 100 m across 433,000 km² of Western Europe — the first system to attain such precision at continental scale.
GeoLink: Empowering Remote Sensing Foundation Model with OpenStreetMap Data: GeoLink directly integrates OpenStreetMap vector data into remote sensing foundation model pretraining by encoding OSM data with a heterogeneous GNN and designing multi-granularity cross-modal learning objectives (region–image-level contrastive + object–patch-level fusion). Pretrained efficiently on 1.27 million sample pairs, GeoLink surpasses existing RS FMs across 7 classification and 4 segmentation/change detection benchmarks.
GreenHyperSpectra: A Multi-Source Hyperspectral Dataset for Global Vegetation Trait Prediction: GreenHyperSpectra constructs a pretraining dataset of 140,000+ multi-source hyperspectral vegetation samples spanning proximal, airborne, and satellite platforms. Label-efficient regression models trained via semi-supervised and self-supervised methods (MAE, GAN, RTM-AE) comprehensively outperform fully supervised baselines on 7 plant trait prediction tasks, with particularly pronounced advantages under label-scarce and out-of-distribution scenarios.
OrbitZoo: Real Orbital Systems Challenges for Reinforcement Learning: This paper presents OrbitZoo, a multi-agent RL environment built on the industrial-grade astrodynamics library Orekit. It integrates high-fidelity orbital dynamics (including atmospheric drag, solar radiation pressure, and third-body effects), a PettingZoo multi-agent interface, and real-time 3D visualization. Validation against real Starlink ephemerides yields a mean MAPE of only 0.16%.
C3PO: Cross-View Cross-Modality Correspondence by Pointmap Prediction: This paper introduces the C3 dataset comprising 90K ground photo–floor plan pairs (597 scenes, 153M pixel-level correspondences, and 85K camera poses), exposes the limitations of existing correspondence models under cross-view cross-modality settings (e.g., ground photos vs. floor plans), and demonstrates that training on this dataset reduces the RMSE of the best-performing baseline by 34%.
Cloud4D: Estimating Cloud Properties at a High Spatial and Temporal Resolution: The first learning framework based on ground-level multi-view cameras that reconstructs four-dimensional (3D spatial + temporal) cloud liquid water content distributions via a homography-guided 2D-to-3D Transformer. The method achieves less than 10% error relative to radar at 25 m spatial and 5 s temporal resolution, improving spatiotemporal resolution by an order of magnitude over satellite observations.
Connecting the Dots: A Machine Learning Dataset for Ionospheric Prediction: This paper constructs an open, ML-ready ionospheric prediction dataset that integrates 8 heterogeneous data sources (solar observations, geomagnetic indices, TEC maps, etc.) spanning approximately 14 years (2010–2024). Three spatiotemporal baseline models—LSTM, SFNO, and GraphCast—are trained on this dataset, achieving TEC forecasts with lead times up to 12 hours.

Browse all 12 Remote Sensing papers →

🧑 Human Understanding (28)¶

KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills: This paper proposes the PBHC framework, which enables a humanoid robot (Unitree G1) to learn highly dynamic whole-body skills such as kung fu and dance through a physics-aware motion processing pipeline and a bi-level optimization scheme for adaptive tracking factors. The approach achieves substantially lower tracking errors than existing methods and is successfully deployed on real hardware.
OmniGaze: Reward-inspired Generalizable Gaze Estimation in the Wild: This paper proposes OmniGaze, a semi-supervised 3D gaze estimation framework that employs a reward model — fusing visual embeddings, MLLM-generated semantic gaze descriptions, and geometric direction vectors — to assess pseudo-label quality. Trained on 1.4 million unlabeled face images, OmniGaze achieves state-of-the-art performance under both within-domain and cross-domain settings across 5 datasets, and demonstrates zero-shot generalization on 4 unseen datasets.
MOSPA: Human Motion Generation Driven by Spatial Audio: This work introduces the novel task of spatial-audio-driven human motion generation, constructs the SAM dataset comprising 9+ hours, 27 scenes, and 12 subjects with paired binaural audio and motion data, and proposes the MOSPA diffusion model. By fusing audio features including MFCC, tempogram, and RMS with sound source position and motion style conditions, MOSPA achieves an FID of 7.98, substantially outperforming music/dance baselines such as EDGE (14.0) and POPDG (21.0).
Learning Dense Hand Contact Estimation from Imbalanced Data: This paper proposes the HACO framework, which addresses class imbalance via Balanced Contact Sampling (BCS) and spatial imbalance via a Vertex-level Class-Balanced Loss (VCB Loss). HACO is the first dense hand contact estimation model trained across 14 datasets (655K images) and achieves state-of-the-art performance across diverse interaction scenarios.
HOI-Dyn: Learning Interaction Dynamics for Human-Object Motion Diffusion: This paper models human-object interaction (HOI) generation as a Driver-Responder system, employing a lightweight Transformer-based interaction dynamics model to explicitly predict how objects respond to human actions. A residual dynamics loss is introduced during training to enforce causal consistency, while inference efficiency is preserved.
ConceptScope: Characterizing Dataset Bias via Disentangled Visual Concepts: This paper proposes ConceptScope, a framework that trains sparse autoencoders (SAE) on representations from visual foundation models to automatically discover and quantify visual concept biases in datasets, categorizing concepts into target / context / bias without any manual annotation.
CPEP: Contrastive Pose-EMG Pre-training Enhances Gesture Generalization on EMG Signals: This paper proposes the CPEP framework, which employs contrastive learning to align low-quality EMG signal representations with high-quality hand pose representations, endowing the EMG encoder with pose-awareness. CPEP is the first to achieve zero-shot recognition of unseen gestures from EMG signals, yielding a 21% improvement on in-distribution gesture classification and a 72% improvement on unseen gesture classification.
RAPTR: Radar-Based 3D Pose Estimation Using Transformer: This paper proposes RAPTR, the first Transformer framework for radar-based 3D human pose estimation using weak supervision (3D bounding boxes + 2D keypoint labels). Through pseudo-3D deformable attention and structured loss functions, RAPTR substantially outperforms baselines on two indoor datasets.
VASA-3D: Lifelike Audio-Driven Gaussian Head Avatars from a Single Image: This paper presents VASA-3D, which adapts VASA-1's 2D motion latent space to a 3D Gaussian splatting representation and leverages VASA-1-synthesized training data for single-image customization, enabling real-time generation (512×512, 75 fps) of lifelike audio-driven 3D head avatars from a single portrait image.
VimoRAG: Video-based Retrieval-augmented 3D Motion Generation for Motion Language Models: This paper proposes VimoRAG, a framework that leverages large-scale in-the-wild video databases as 2D motion priors to enhance 3D motion generation. Two core bottlenecks—human motion video retrieval and error propagation—are addressed via the Gemini-MVR retriever and the McDPO training strategy.

Browse all 28 Human Understanding papers →

📹 Video Understanding (41)¶

InfiniPot-V: Memory-Constrained KV Cache Compression for Streaming Video Understanding: This paper proposes InfiniPot-V, the first training-free and query-agnostic streaming video understanding framework. It achieves online KV cache compression via two complementary metrics — Temporal-axis Redundancy (TaR) and Value-Norm (VaN) — enabling streaming video understanding of arbitrary length under a fixed memory budget.
VGEnt: Graph-Based Retrieval-Reasoning-Augmented Generation for Long Video Understanding: This paper proposes VGEnt, a graph-based retrieval-reasoning-augmented generation framework that constructs a video knowledge graph to preserve cross-segment semantic relationships, and introduces structured reasoning steps to filter noise and aggregate information. VGEnt consistently improves open-source LVLMs by 3.0%–5.4% across multiple long video understanding benchmarks and outperforms existing video RAG methods by 8.6%.
Seeing the Arrow of Time in Large Multimodal Models: This paper reveals that current large multimodal models (LMMs) are surprisingly insensitive to the temporal directionality of video (i.e., the Arrow of Time)—producing nearly identical answers for forward and reversed playback. The authors propose ArrowRL, a GRPO-based training strategy that introduces a reverse video reward to elicit temporal direction awareness, and construct AoTBench for evaluation. The approach achieves significant gains across multiple VQA benchmarks, including a 65.9% relative improvement on Vinoground.
VADTree: Explainable Training-Free Video Anomaly Detection via Hierarchical Granularity: This paper proposes VADTree, a training-free video anomaly detection framework that leverages a pretrained Generic Event Boundary Detection (GEBD) model to construct a Hierarchical Granularity-aware Tree (HGTree), enabling adaptive sampling and multi-granularity reasoning for anomalous events of varying temporal spans. VADTree achieves state-of-the-art performance among training-free methods on three benchmarks—UCF-Crime, XD-Violence, and MSAD—and even surpasses certain weakly supervised approaches.
PANDA: Towards Generalist Video Anomaly Detection via Agentic AI Engineer: This paper proposes PANDA, an agentic AI engineer framework built upon MLLMs, which achieves training-free and human-intervention-free generalist video anomaly detection through four core capabilities: adaptive scene-aware strategy planning, goal-driven heuristic reasoning, tool-augmented self-reflection, and chain-of-memory.
Unleashing Hour-Scale Video Training for Long Video-Language Understanding: This work constructs VideoMarathon, the first large-scale hour-level video instruction-following dataset (9,700 hours, 3.3M QA pairs, 22 task types), and proposes Hour-LLaVA, a model that leverages a memory repository, forgetting mechanism, and MemAug module to enable efficient training and inference on hour-scale videos at 1 FPS, achieving state-of-the-art results among open-source models of comparable scale across four long video benchmarks.
VideoLucy: Deep Memory Backtracking for Long Video Understanding: This paper proposes VideoLucy, a framework that simulates the human coarse-to-fine recall process via a hierarchical memory structure and an agent-based iterative backtracking mechanism. VideoLucy substantially outperforms existing methods on multiple long video understanding benchmarks, surpassing even commercial models such as GPT-4o.
EgoGazeVQA: Egocentric Gaze-Guided Video Question Answering Benchmark: This paper introduces EgoGazeVQA, the first egocentric video question answering benchmark that incorporates user eye-gaze data. Through gaze-guided prompting strategies (textual, visual, and salience map), the benchmark demonstrates substantial improvements in MLLMs' ability to understand user intent. The Gaze Salience Map strategy raises MiniCPM-o's accuracy from 35.9% to 53.7%.
TempSamp-R1: Effective Temporal Sampling with Reinforcement Fine-Tuning for Video LLMs: TempSamp-R1 is a reinforcement fine-tuning framework that addresses the inefficiency of on-policy sampling in GRPO for video temporal grounding—caused by the vast search space—by introducing ground-truth annotations as off-policy supervision signals, non-linear soft advantage estimation, and a hybrid CoT training paradigm, achieving new state-of-the-art results on Charades-STA, ActivityNet, and QVHighlights.
MoniTor: Exploiting Large Language Models with Instruction for Online Video Anomaly Detection: This paper proposes MoniTor, a memory-based online scoring queue framework that leverages LLMs for training-free online video anomaly detection (VAD). It guides LLMs toward real-time anomaly recognition through a dual-layer memory mechanism, behavior prediction, and a standard scoring queue.

Browse all 41 Video Understanding papers →

🚗 Autonomous Driving (47)¶

FutureSightDrive: Thinking Visually with Spatio-Temporal CoT for Autonomous Driving: FSDrive enables VLAs to "think visually" — first acting as a world model to generate a unified visual CoT frame that integrates future lane lines, 3D detection boxes, and scene predictions, then acting as an inverse dynamics model to perform trajectory planning based on current observations and the visual CoT. This approach activates the visual generation capability of MLLMs using only a minimal amount of data (~0.3%).
StreamForest: Efficient Online Video Understanding with Persistent Event Memory: This paper proposes StreamForest, an architecture that adaptively organizes streaming video frames into multiple event-level tree structures via a "Persistent Event Memory Forest," combined with a "Fine-grained Spatiotemporal Window" to capture short-term visual cues. The method achieves 77.3% accuracy on StreamingBench and retains 96.8% of performance under extreme compression (only 1024 visual tokens).
RAW2Drive: Reinforcement Learning with Aligned World Models for End-to-End Autonomous Driving: This paper proposes RAW2Drive, the first model-based reinforcement learning (MBRL) end-to-end autonomous driving framework operating directly from raw sensor inputs to planning. Through a dual-stream world model design — first training a privileged world model, then guiding a raw-sensor world model via an alignment mechanism — RAW2Drive achieves state-of-the-art performance on CARLA v2 and Bench2Drive, substantially outperforming imitation learning (IL) methods.
DBLoss: Decomposition-based Loss Function for Time Series Forecasting: This paper proposes DBLoss—a general-purpose loss function based on exponential moving average (EMA) decomposition. During loss computation, both predictions and ground-truth values are decomposed into seasonal and trend components within the forecasting horizon, and losses are computed separately for each component. DBLoss serves as a plug-and-play replacement for MSE and consistently improves any deep learning forecasting model, with effectiveness validated across 8 benchmark datasets × 8 SOTA models.
DINO-Foresight: Looking into the Future with DINO: This paper proposes DINO-Foresight, which forecasts future-frame feature evolution within the semantic feature space of a Vision Foundation Model (VFM). A self-supervised Masked Feature Transformer predicts PCA-compressed representations of multi-layer DINOv2 features. Paired with plug-and-play task-specific heads, a single model simultaneously handles semantic segmentation, instance segmentation, depth estimation, and surface normal prediction, substantially outperforming the VISTA world model while achieving 100× faster inference.
DriveDPO: Policy Learning via Safety DPO For End-to-End Autonomous Driving: DriveDPO is a two-stage framework that first fuses human-imitation similarity and rule-based safety scores into a single supervised distribution via unified policy distillation, then applies Safety DPO to construct trajectory preference pairs of the form "human-like but unsafe vs. human-like and safe" for policy fine-tuning — achieving a new state-of-the-art PDMS of 90.0 on NAVSIM.
V2X-Radar: A Multi-Modal Dataset with 4D Radar for Cooperative Perception: This paper presents V2X-Radar, the first large-scale real-world multi-modal vehicle-to-everything (V2X) cooperative perception dataset incorporating 4D radar, LiDAR, and multi-view camera data. The dataset covers diverse weather and lighting conditions, providing 20K LiDAR frames, 40K camera images, 20K 4D radar scans, and 350K annotated bounding boxes, along with comprehensive benchmarks across three sub-datasets.
Flow Matching-Based Autonomous Driving Planning with Advanced Interactive Behavior Modeling: This paper proposes Flow Planner—a system combining three synergistic innovations: fine-grained trajectory tokenization, an interaction-enhanced spatiotemporal fusion architecture, and flow matching with classifier-free guidance. It is the first purely learning-based method to surpass 90 points on nuPlan Val14 (90.43), and outperforms Diffusion Planner by 8.92 points on the interaction-intensive interPlan benchmark.
Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution: Proposes SeerDrive, which achieves SOTA on NAVSIM and nuScenes through bidirectional modeling of scene evolution and trajectory planning (future-aware planning + iterative interaction).
Unifying Appearance Codes and Bilateral Grids for Driving Scene Gaussian Splatting: A multi-scale bilateral grid pyramid is proposed to unify global appearance codes and pixel-level bilateral grids. A 3-level hierarchy (coarse→medium→fine) captures global/regional/pixel-level photometric variation respectively. By employing a luminance-guided slice-and-blend pipeline and adaptive regularization, the method addresses photometric inconsistency in driving scene 3DGS, achieving a 28.2% improvement in Chamfer Distance over OmniRe on Waymo.

Browse all 47 Autonomous Driving papers →

🤖 Robotics & Embodied AI (80)¶

AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning: AutoVLA integrates physical action tokens directly into a pretrained VLM (Qwen2.5-VL-3B), equips the model with fast/slow dual-thinking modes via SFT, and applies GRPO reinforcement fine-tuning to enable adaptive reasoning switching and optimize planning performance. The approach achieves competitive end-to-end driving performance across four major autonomous driving benchmarks: nuPlan, Waymo, nuScenes, and CARLA.
ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning: ThinkAct proposes a dual-system framework that applies action-aligned visual rewards to fine-tune MLLMs via reinforcement learning, eliciting embodied reasoning capabilities and compressing reasoning plans into visual latent representations to guide a downstream action model—realizing a "think before act" VLA reasoning paradigm.
Reinforcement Learning with Action Chunking: This paper proposes Q-chunking, which extends action chunking from imitation learning to TD-based reinforcement learning by running RL directly over a "chunked" action space, thereby improving exploration and sample efficiency in long-horizon sparse-reward tasks.
SAFE: Multitask Failure Detection for Vision-Language-Action Models: SAFE identifies consistent "failure regions" in the internal feature space of VLA models that generalize across tasks. Leveraging this observation, it trains lightweight MLP/LSTM failure detectors and applies Functional Conformal Prediction (FCP) for threshold calibration. The approach achieves 78% ROC-AUC on unseen tasks with less than 1% computational overhead, substantially outperforming token-uncertainty and action-consistency baselines.
SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning: This work is the first to systematically apply the Constrained Markov Decision Process (CMDP) framework from Safe Reinforcement Learning (SafeRL) to safety alignment of Vision-Language-Action (VLA) models. Through a four-stage Integrated Safety Approach (ISA)—Model, Elicit, Constrain, and Assure—the method achieves an 83.58% reduction in safety violation costs on mobile manipulation tasks while maintaining task performance (+3.85%).
EgoBridge: Domain Adaptation for Generalizable Imitation from Egocentric Human Data: This paper proposes EgoBridge, a framework that uses Optimal Transport (OT) to align the joint distribution (features + actions) of human and robot data in a shared policy latent space, combined with Dynamic Time Warping (DTW) to construct pseudo-pairs, enabling cross-embodiment knowledge transfer from egocentric human data to robots, achieving up to 44% absolute improvement in success rate on real-world tasks.
Self-Improving Embodied Foundation Models: This paper proposes a two-stage post-training framework for embodied foundation models: Stage 1 performs supervised fine-tuning via behavior cloning and steps-to-go prediction; Stage 2 leverages the resulting self-reward function and success detector for online RL self-improvement. Using only 1–3% additional data, the method achieves over 1.5× improvement in success rate and, for the first time, demonstrates a robot autonomously acquiring novel skills beyond the distribution of imitation data.
RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation: This paper proposes RoboCerebra, a long-horizon robotic manipulation benchmark comprising 1,000 human demonstration trajectories (averaging 2,972 steps, approximately 6× longer than existing benchmarks). Through a hierarchical planning and execution framework and a multi-dimensional evaluation protocol, it systematically assesses VLMs across three System 2 cognitive dimensions: planning, reflection, and memory.
BEAST: Efficient Tokenization of B-Splines Encoded Action Sequences for Imitation Learning: BEAST parameterizes action sequences via B-splines—estimating control points through ridge regression and uniformly quantizing them into fixed-length tokens—achieving 20× token compression (100 steps → 5 tokens), mathematically guaranteed $C^0$ continuity across action chunks, a top-1 success rate on LIBERO-Long (86.4%), and an inference throughput of 617 Hz (2.14× faster than π₀ and 101× faster than OpenVLA).
Spatial Understanding from Videos: Structured Prompts Meet Simulation Data: This paper proposes a two-pronged approach combining the SpatialMind structured prompting strategy and the ScanForgeQA synthetic QA dataset to substantially enhance VLMs' ability to perform 3D spatial reasoning from scanned videos, without modifying the underlying model architecture.

Browse all 80 Robotics & Embodied AI papers →

🎮 Reinforcement Learning (143)¶

Training Language Models to Reason Efficiently: By incorporating a length penalty term into the RL reward—multiplying the correctness reward by $(1 - \alpha \cdot \sigma(\text{norm\_len}))$—and using a single hyperparameter $\alpha$ to control the token–accuracy trade-off curve, this work achieves a 50% reduction in token usage with less than 5% accuracy degradation on 7B reasoning models after only 100 RL training steps.
MMaDA: Multimodal Large Diffusion Language Models: This paper presents MMaDA, the first multimodal foundation model that simultaneously achieves text reasoning, multimodal understanding, and text-to-image generation within a unified discrete diffusion architecture. MMaDA bridges the gap between diffusion model pre-training and post-training through mixed long chain-of-thought (CoT) fine-tuning and the UniGRPO reinforcement learning algorithm.
Generalizing Verifiable Instruction Following: This paper introduces IFBench, a benchmark for evaluating generalization in precise instruction following, demonstrating that current SOTA models severely overfit to the 25 constraint templates of IFEval. It further proposes IF-RLVR, a training method based on GRPO with verifiable rewards, which significantly improves both in-domain and out-of-domain instruction following performance.
Reasoning Gym: Reasoning Environments for Reinforcement Learning with Verifiable Rewards: This work releases Reasoning Gym, a library of 100+ procedurally generated reasoning tasks spanning algebra, arithmetic, algorithms, logic, geometry, graph theory, games, and more. Each task supports infinite data generation and parameterized difficulty control. Experiments demonstrate that RLVR training achieves significant skill transfer both within and across domains, and improves performance on external benchmarks such as MATH and GSM8K.
Checklists Are Better Than Reward Models For Aligning Language Models: This paper proposes Reinforcement Learning from Checklist Feedback (RLCF), which decomposes instructions into dynamically generated yes/no checklists, scores each item using an AI judge and code verifier, and trains with DPO. RLCF consistently improves Qwen2.5-7B-Instruct across 5 benchmarks and is the only method that achieves positive gains on all benchmarks (FollowBench +4pt, InFoBench +6pt, Arena-Hard +3pt).
Note 5: ReSearch — Learning to Reason with Search: ReSearch embeds search operations as first-class primitives within reasoning chains and leverages GRPO reinforcement learning to automatically learn when and how to search—without any supervision on intermediate reasoning steps—achieving an average relative improvement of 15.81% over baselines on multi-hop QA benchmarks.
NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation: This paper proposes NoisyRollout, a data augmentation method with zero additional training cost. During GRPO-based VLM training, it mixes rollouts from clean and moderately perturbed images to enhance policy exploration diversity. Using only 2.1K samples, it achieves state-of-the-art performance among open-source RL fine-tuned models across five out-of-domain benchmarks.
Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning: Router-R1 frames multi-LLM routing and aggregation as a sequential decision-making process, employing an LLM itself as the router to interleave think and route actions. Trained via PPO with a triple reward covering format, correctness, and cost, Router-R1 outperforms all router baselines across 7 QA benchmarks and generalizes to previously unseen LLMs.
Horizon Reduction Makes RL Scalable: Through large-scale experiments involving up to one billion transitions, this paper identifies the curse of horizon—excessively long decision horizons—as the primary scalability bottleneck in offline RL, and demonstrates that horizon reduction techniques such as n-step returns and hierarchical policies substantially improve scalability. Building on this analysis, the paper proposes SHARSA, a simple yet effective method.
Meta-World+: An Improved, Standardized, RL Benchmark: This paper systematically exposes how undocumented reward function inconsistencies across versions of the Meta-World benchmark distort algorithm comparisons, and releases a standardized new version, Meta-World+, which explicitly retains both V1 and V2 reward functions, introduces MT25/ML25 task sets, upgrades to the Gymnasium API, and enables fully reproducible evaluation for multi-task and meta-reinforcement learning.

Browse all 143 Reinforcement Learning papers →

🎁 Recommender Systems (24)¶

Measuring What Matters: Construct Validity in Large Language Model Benchmarks: This paper presents a systematic review of 445 LLM benchmark papers conducted by 29 experts, examining existing LLM evaluation benchmarks through the lens of construct validity across four dimensions — phenomenon definition, task design, scoring metrics, and conclusion claims — and proposes 8 actionable recommendations for improvement.
Inference-Time Reward Hacking in Large Language Models: This paper mathematically proves that inference-time alignment methods (e.g., BoN) inevitably exhibit reward hacking (true reward first increases then decreases) when optimizing a proxy reward. It proposes Best-of-Poisson (BoP) sampling to approximate the optimal KL-reward trade-off distribution, and designs the HedgeTune algorithm to locate the optimal inference-time parameter via one-dimensional root-finding, effectively mitigating reward hacking in both mathematical reasoning and human preference settings.
Overcoming Sparsity Artifacts in Crosscoders to Interpret Chat-Tuning: This work identifies two categories of sparsity artifacts introduced by L1 loss in Crosscoders—Complete Shrinkage (which erroneously zeros out weakly shared concepts) and Latent Decoupling (which splits shared concepts into spurious model-specific latents)—and proposes Latent Scaling as a diagnostic tool and BatchTopK Crosscoder as an alternative training scheme, substantially improving the reliability of chat-tuning concept discovery.
Validating LLM-as-a-Judge Systems under Rating Indeterminacy: This paper proposes a framework for validating LLM-as-a-Judge systems under rating indeterminacy, replacing forced-choice rating with a "response set" multi-label rating scheme, achieving up to 31% performance improvement in the selected judge system.
Transformer Copilot: Learning from The Mistake Log in LLM Fine-tuning: This paper proposes the Transformer Copilot framework, which systematically records a "Mistake Log" during LLM fine-tuning, trains an auxiliary Copilot model to learn the Pilot's error patterns, and rectifies logits at inference time to improve generation quality, achieving up to 34.5% improvement across 12 benchmarks.
R²ec: Towards Large Recommender Models with Reasoning: This paper proposes R²ec, the first unified large recommender model that endogenously integrates reasoning capabilities, achieving joint reasoning chain generation and efficient item prediction via a dual-head architecture, and introduces the RecPO reinforcement learning framework to jointly optimize reasoning and recommendation objectives without any annotated reasoning data.
Semantic Retrieval Augmented Contrastive Learning for Sequential Recommendation: This paper proposes SRA-CL, a framework that leverages the semantic understanding capabilities of LLMs to construct high-quality contrastive sample pairs. By combining semantic retrieval with a learnable sample synthesizer, SRA-CL enhances contrastive learning for sequential recommendation and achieves state-of-the-art performance across four datasets in a plug-and-play manner.
Think before Recommendation: Autonomous Reasoning-enhanced Recommender: This paper proposes RecZero (a pure RL paradigm) and RecOne (a hybrid SFT+RL paradigm), abandoning conventional teacher-student distillation. Both approaches leverage GRPO-based reinforcement learning to train a single LLM to autonomously develop reasoning capabilities for rating prediction. A structured "Think-before-Recommendation" template guides step-by-step reasoning (user analysis → item analysis → matching → rating), achieving significant improvements over existing baselines across four datasets.
Position: Towards Bidirectional Human-AI Alignment: This paper proposes a Bidirectional Human-AI Alignment framework grounded in a systematic review of 400+ papers, arguing that AI alignment should not be limited to the unidirectional goal of "aligning AI with humans," but must also encompass the critically underexplored direction of "aligning humans with AI," while identifying key gaps in the current research landscape.
MMPB: It's Time for Multi-Modal Personalization: This paper introduces MMPB, the first VLM personalization evaluation benchmark, comprising 111 personalizable concepts, 10k+ image-text QA pairs, and 15 task types. Evaluation of 23 VLMs reveals that even the strongest model, GPT-4o, performs poorly on personalization tasks, exposing critical limitations in preference reasoning, visual cue utilization, and conflicts between safety alignment and personalization.

Browse all 24 Recommender Systems papers →

🔄 Self-Supervised Learning (37)¶

TabArena: A Living Benchmark for Machine Learning on Tabular Data: This paper introduces TabArena, the first continuously maintained "living" benchmark for tabular machine learning. From 1,053 candidate datasets, 51 are curated and 16 models are evaluated through large-scale experiments (~25 million model training runs). Key findings: under post-hoc ensembling, deep learning models match or surpass GBDTs; tabular foundation models excel on small datasets; and cross-model ensembles further advance the state of the art.
Mitra: Mixed Synthetic Priors for Enhancing Tabular Foundation Models: This paper presents the first systematic study of design principles for synthetic priors, identifying diversity, distinctiveness, and real-data alignment as critical attributes. Based on these findings, the authors propose Mitra — a tabular foundation model trained on a carefully selected mixture of synthetic priors — which consistently outperforms TabPFNv2 and TabICL on both classification and regression benchmarks.
MMTU: A Massive Multi-Task Table Understanding and Reasoning Benchmark: This paper introduces MMTU, a large-scale benchmark comprising 28,136 questions spanning 25 real-world table tasks, designed to systematically evaluate LLMs on professional-level table understanding, reasoning, and manipulation. Even frontier reasoning models such as GPT-5 achieve only approximately 69.6% on this benchmark.
TabSTAR: A Tabular Foundation Model for Tabular Data with Text Fields: TabSTAR is a foundation model designed specifically for tabular data with text fields. It achieves target-aware text representations through end-to-end optimization with an unfrozen text encoder (e5-small-v2), injects target semantics via target-aware tokens, and enables cross-dataset transfer learning through a dataset-parameter-free architecture. After pre-training on 350 datasets, TabSTAR surpasses CatBoost-Tuned (4h tuning) on 12 out of 14 classification datasets and outperforms TabPFN-v2 on 8 out of 11 datasets.
DataRater: Meta-Learned Dataset Curation: This paper proposes DataRater, a meta-gradient-based data valuation framework that employs meta-learning to automatically score and filter low-quality training samples. It achieves up to 46.6% net compute savings across multiple pre-training datasets, and a DataRater trained on a 400M internal model generalizes directly to LLM training at scales ranging from 50M to 1B parameters.
Continuous Subspace Optimization for Continual Learning (CoSO): This paper proposes CoSO, a framework that dynamically derives continuous subspaces from per-step gradient SVD (rather than LoRA's fixed subspace), combined with orthogonal projection onto historical task subspaces to prevent interference and Frequent Directions for efficient gradient information aggregation. CoSO achieves 78.19% final accuracy on ImageNet-R with 20 tasks, surpassing the best baseline by 2.77 percentage points.
SEAL: Semantic-Aware Hierarchical Learning for Generalized Category Discovery: This paper proposes SEAL, a framework that leverages naturally occurring semantic hierarchies (rather than manually constructed abstract hierarchies) to guide generalized category discovery. Through hierarchically semantic-guided soft contrastive learning and a cross-granularity consistency module, SEAL achieves state-of-the-art performance on fine-grained benchmarks.
Contrastive Representations for Temporal Reasoning: This paper proposes CRTR (Contrastive Representations for Temporal Reasoning), which introduces intra-trajectory negative pairs by repeating trajectory IDs within training batches. This eliminates the reliance on static contextual features in standard temporal contrastive learning, enabling representations that reflect temporal structure. CRTR achieves, for the first time, search-free solving on combinatorial reasoning tasks such as the Rubik's Cube.
Connecting Jensen-Shannon and Kullback-Leibler Divergences: A New Bound for Representation Learning: This paper derives the optimal tight lower bound of KL divergence in terms of JS divergence, $\Xi(D_{\text{JS}}) \leq D_{\text{KL}}$, in the general case. It proves that training a discriminator by minimizing cross-entropy loss is equivalent to maximizing a guaranteed lower bound on mutual information, thereby providing the missing theoretical foundation for JSD-based discriminative representation learning methods. The tightness and practical utility of the bound are validated in MI estimation and the Information Bottleneck framework.
Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning: This paper theoretically proves that self-supervised contrastive learning (DCL) is approximately equivalent to a supervised contrastive loss (NSCL), with the gap vanishing at rate $O(1/C)$ as the number of classes increases. It further proves that the global optimum of NSCL satisfies Neural Collapse (augmentation collapse + within-class collapse + Simplex ETF), and proposes a tighter few-shot error bound based on directional CDNV.

Browse all 37 Self-Supervised Learning papers →

📐 Optimization & Theory (127)¶

Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry: This paper reveals a fundamental duality between sparse autoencoder (SAE) architectures and the concept structures they are capable of discovering — each SAE implicitly assumes a particular organization of concepts, and when this assumption is mismatched, concepts are systematically missed. Based on this analysis, the authors propose SpaDE, a novel SAE that accounts for nonlinear separability and dimensional heterogeneity.
Small Batch Size Training for Language Models: When Vanilla SGD Works, and Why Gradient Accumulation Is Wasteful: This paper systematically investigates the behavior of small batch sizes (down to batch size = 1) in language model pre-training and fine-tuning. It proposes a scaling rule for Adam $\beta_2$ based on fixing the "token half-life," demonstrates that small-batch training is stable, shows that vanilla SGD becomes competitive with adaptive optimizers under small batches, and recommends avoiding gradient accumulation.
In Search of Adam's Secret Sauce: Through large-scale experiments training 1500+ language models, this paper establishes: (1) Signum closes 96% of the SGD–Adam gap yet remains 25% slower than Adam; (2) setting $\beta_1 = \beta_2$ is a near-optimal simplification of Adam; (3) under $\beta_1 = \beta_2 = \beta$, Adam can be reinterpreted as a signal-to-noise ratio–adaptive Signum that estimates the gradient mean and variance via online Gaussian variational inference.
FlyLoRA: Boosting Task Decoupling and Parameter Efficiency via Implicit Rank-Wise Mixture-of-Experts: Inspired by the fly olfactory circuit, FlyLoRA replaces the down-projection matrix $A$ in LoRA with a frozen sparse random projection and employs top-$k$ activation selection to realize implicit rank-wise MoE routing. This design eliminates routing parameters, reduces intra-task interference, and naturally supports multi-task model merging by exploiting the near-orthogonality of random projections.
MergeBench: A Benchmark for Merging Domain-Specialized LLMs: MergeBench is the first comprehensive benchmark suite for evaluating large-scale domain-specialized LLM merging, covering Llama and Gemma families up to 9B parameters, five task domains, and eight merging methods, providing systematic evaluation and practical guidelines across three dimensions: multi-task performance, forgetting, and runtime efficiency.
Emergence and Scaling Laws in SGD Learning of Shallow Neural Networks: This paper provides a precise analysis of online SGD learning of additive models (sums of single-index functions) in shallow neural networks, proving that the learning of each teacher neuron exhibits a sharp phase transition (emergence), and that the superposition of many such transition curves across different timescales naturally produces a smooth power-law scaling law.
MDNS: Masked Diffusion Neural Sampler via Stochastic Optimal Control: This paper proposes the Masked Diffusion Neural Sampler (MDNS), a framework grounded in stochastic optimal control theory for continuous-time Markov chains (CTMCs). By aligning path measures, MDNS trains a discrete neural sampler capable of accurately sampling from Ising/Potts models with state spaces as large as $10^{122}$, substantially outperforming existing learning-based baselines.
Purifying Shampoo: Investigating Shampoo's Heuristics by Decomposing its Preconditioner: By decomposing Shampoo's preconditioner into eigenvalue and eigenbasis components, this work reveals that learning rate grafting essentially compensates for eigenvalue staleness and scaling bias, and proposes eigenvalue correction and adaptive eigenbasis update frequency as principled replacements for these heuristics.
Learning Quadratic Neural Networks in High Dimensions: SGD Dynamics and Scaling Laws: This paper provides a precise analysis of gradient-based training of two-layer neural networks with quadratic activations in high dimensions. Under the setting where data is generated by $f_*(x) \propto \sum_{j=1}^r \lambda_j \sigma(\langle \theta_j, x \rangle)$, with extensive width $r \asymp d^\beta$ and power-law decaying coefficients $\lambda_j \asymp j^{-\alpha}$, the paper derives scaling laws for the prediction risk as a function of optimization time, sample size, and model width.
Conformal Prediction for Causal Effects of Continuous Treatments: This work is the first to construct conformal prediction intervals for causal effects of continuous treatments (e.g., drug dosage) by parameterizing intervention-induced propensity shifts via a tilting function family and employing quantile regression, providing finite-sample $1-\alpha$ coverage guarantees under both known and unknown propensity score settings.

Browse all 127 Optimization & Theory papers →

📐 Learning Theory (25)¶

The Structural Complexity of Matrix-Vector Multiplication: This paper proves that for Boolean matrices $\mathbf{M} \in \{0,1\}^{m \times n}$ with corrupted VC-dimension $d$, matrix-vector multiplication can be performed in $\widetilde{O}(nm^{1-1/d}+m)$ time. This is the first truly sub-quadratic upper bound for structured matrices, refuting the applicability of the OMv conjecture on structured inputs, and yields the first high-accuracy sub-quadratic algorithms for dynamic Laplacian solving, effective resistance, triangle detection, and related problems.
Non-Clairvoyant Scheduling with Progress Bars: This paper introduces a "progress bar" information model as an interpolation framework between clairvoyant and non-clairvoyant scheduling. It designs scheduling algorithms with optimal consistency–robustness tradeoffs for both adversarial and stochastic progress bars, while advancing the theoretical frontier of learning-augmented scheduling.
Optimism Without Regularization: Constant Regret in Zero-Sum Games: This paper provides the first proof that Optimistic Fictitious Play without regularization achieves $O(1)$ constant regret in $2\times2$ zero-sum games, matching the optimal rate of regularized Optimistic FTRL. It further establishes an $\Omega(\sqrt{T})$ regret lower bound for Alternating Fictitious Play, separating the capabilities of optimism and alternation in the unregularized setting.
Improved Approximation Algorithms for Chromatic and Pseudometric-Weighted Correlation Clustering: For two important generalizations of Correlation Clustering—Chromatic CC and pseudometric-weighted CC—this paper achieves a 2.15-approximation and a tight 10/3-approximation, respectively, via LP relaxation and carefully designed rounding functions, significantly improving upon the previous best results of 2.5 and 6.
On Agnostic PAC Learning in the Small Error Regime: In the small error regime of agnostic PAC learning ($\tau \approx d/m$), this paper constructs a computationally efficient learner based on ERM aggregation that achieves an error upper bound of $c \cdot \tau + O(\sqrt{\tau d/m} + d/m)$ with $c \leq 2.1$, matching known lower bounds and advancing the precise complexity characterization of agnostic learning.
Prediction-Powered Semi-Supervised Learning with Online Power Tuning: This paper extends the Prediction-Powered Inference (PPI) framework to the training phase of semi-supervised learning. It proposes an unbiased gradient estimator and designs an online AdaGrad algorithm to dynamically tune the interpolation parameter $\lambda$ between pseudo-labels and true labels, achieving convergence rates matching the optimal fixed $\lambda$ while maintaining unbiasedness.
Finite-Time Analysis of Stochastic Nonconvex Nonsmooth Optimization on the Riemannian Manifolds: This paper proposes the Riemannian Online to NonConvex (RO2NC) algorithm and its zeroth-order variant ZO-RO2NC, establishing for the first time a finite-time sample complexity guarantee of $O(\delta^{-1}\epsilon^{-3})$ for fully nonsmooth nonconvex stochastic optimization on Riemannian manifolds, matching the optimal result in Euclidean space.
Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning: This paper proposes Controllable Pseudo-label Generation (CPG), a framework that progressively incorporates reliable pseudo-labels into the labeled set via a controllable self-reinforcing optimization cycle. By training a Bayes-optimal classifier on a distribution of known composition, CPG achieves accuracy gains of up to 15.97% in the Realistic LTSSL setting where the unlabeled data distribution is entirely unknown.
Conformal Online Learning of Deep Koopman Linear Embeddings: This paper proposes the COLoKe framework, which reinterprets conformal prediction as a model consistency diagnostic tool. Parameter updates are triggered only when the Koopman model's prediction error exceeds a dynamically calibrated threshold, enabling efficient online Koopman linear embedding learning for nonlinear dynamical systems.
How Many Domains Suffice for Domain Generalization? A Tight Characterization via the Domain Shattering Dimension: This paper introduces the Domain Shattering Dimension (Gdim), a novel combinatorial measure that tightly characterizes the number of domains required for domain generalization (i.e., the domain sample complexity), and establishes its relationship to the classical VC dimension as $\Theta(d \log(1/\alpha))$.

Browse all 25 Learning Theory papers →

🔗 Causal Inference (20)¶

Do-PFN: In-Context Learning for Causal Effect Estimation: This paper proposes Do-PFN, which extends Prior-data Fitted Networks (PFN) to causal effect estimation. A Transformer is pre-trained on large-scale synthetic SCM data to perform in-context causal reasoning, enabling prediction of causal intervention distributions (CID) and CATE from observational data alone—without requiring causal graph knowledge or the unconfoundedness assumption—achieving strong performance on both synthetic and semi-synthetic benchmarks.
Performative Validity of Recourse Explanations: This paper formally analyzes the "performative" effects of recourse explanations — when a large number of rejected applicants act on recourse recommendations, their collective behavior induces distribution shift that renders recourse invalid after model retraining — and proves that only Improvement-based Causal Recourse (ICR), which intervenes solely on causal variables, preserves "performative validity" under broad conditions.
Counterfactual Reasoning for Steerable Pluralistic Value Alignment of Large Language Models: This paper proposes the COUPLE framework, which constructs a Structural Causal Model (SCM) to model the dependencies and priorities among multi-dimensional values, and leverages counterfactual reasoning to achieve steerable alignment of LLMs toward arbitrary fine-grained pluralistic value objectives.
It's Hard to Be Normal: The Impact of Noise on Structure-agnostic Estimation: This paper proves that Double Machine Learning (DML) is minimax optimal under Gaussian treatment noise ($O(\epsilon^2 + n^{-1/2})$), but becomes suboptimal under non-Gaussian noise. It proposes Agnostic Cumulant-based Estimation (ACE), which exploits higher-order cumulants to achieve $r$-th order insensitivity $O(\epsilon^r + n^{-1/2})$.
LLM Interpretability with Identifiable Temporal-Instantaneous Representation: This paper proposes an identifiable temporal causal representation learning framework for the high-dimensional activation spaces of LLMs. By adopting a linearized formulation that jointly models time-lagged and instantaneous causal relationships, it resolves the computational bottleneck that prevents existing CRL methods from scaling to LLM-scale dimensions, while preserving theoretical identifiability guarantees.
Practical do-Shapley Explanations with Estimand-Agnostic Causal Inference: This paper proposes the Estimand-Agnostic (EA) approach and the Frontier-Reducibility Algorithm (FRA) for efficient computation of causal Shapley values (do-SV). By training a single SCM to learn the observational distribution, the framework answers arbitrary identifiable causal queries and reduces the number of coalitions requiring evaluation by approximately 90% via coalition reduction.
Demystifying Spectral Feature Learning for Instrumental Variable Regression: This paper establishes rigorous generalization error bounds for spectral feature-based nonparametric instrumental variable (NPIV) regression, revealing that performance is jointly governed by two factors: spectral alignment between the structural function and the conditional expectation operator (approximation error) and the rate of singular value decay (estimation error). A Good-Bad-Ugly trichotomy is proposed along with data-driven diagnostic tools.
Revealing Multimodal Causality with Large Language Models: This paper proposes MLLM-CD, the first framework for causal discovery from multimodal unstructured data (text + images). It identifies causal variables via contrastive factor discovery, infers causal structure through statistical methods, and resolves structural ambiguity via iterative multimodal counterfactual reasoning.
Causality-Induced Positional Encoding for Transformer-Based Representation Learning of Non-Sequential Features: CAPE learns the causal DAG structure among features from tabular data, embeds it into hyperbolic space to generate causality-aware rotary positional encodings (RoPE), enabling Transformers to process non-sequential yet causally structured feature data, with significant performance gains on downstream multi-omics tasks.
Characterization and Learning of Causal Graphs from Hard Interventions: This work systematically analyzes the theoretical advantages of hard interventions in causal discovery with latent variables for the first time. It proposes a generalized do-calculus (4 rules) and a twin augmented MAG graph representation, establishes the necessary and sufficient graphical conditions for $\mathcal{I}$-Markov equivalence classes, and designs a provably sound FCI-variant learning algorithm. Experiments show that hard interventions reduce the equivalence class size by 37-57% compared to soft interventions.

Browse all 20 Causal Inference papers →

🔬 Interpretability (81)¶

A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders: This paper identifies and systematically studies the phenomenon of "feature absorption" in SAEs: apparently monosemantic SAE latents fail to activate on certain tokens because their feature directions are "absorbed" by more specific sub-latents. This is shown to be an inevitable consequence of hierarchical features combined with sparsity loss, posing a fundamental challenge to using SAEs for reliable LLM interpretation.
Reasoning by Superposition: A Theoretical Perspective on Chain of Continuous Thought: This paper theoretically demonstrates the expressive advantage of continuous chain-of-thought (Coconut) on directed graph reachability: a two-layer Transformer using $D$ continuous thought steps suffices to solve graph reachability with diameter $D$, whereas discrete CoT requires $O(n^2)$ steps. The core mechanism is that continuous thought vectors encode multiple search frontiers simultaneously in a "superposition state," enabling implicit parallel BFS.
From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit: Proposes MP-SAE, which unrolls the classic Matching Pursuit algorithm into a sequential encoder for SAEs. By achieving conditional orthogonality through residual-guided greedy feature selection, it captures hierarchical structures, non-linearly accessible features, and cross-modal features that standard SAEs fail to detect, while natively supporting adaptive sparsity adjustment at inference time.
Superposition Yields Robust Neural Scaling: This paper identifies representational superposition as the core driver of neural scaling laws: in the strong-superposition regime, loss universally scales inversely with model dimension ($L \propto 1/m$), independent of the specific form of the data frequency distribution—consistent with empirical scaling behavior in real LLMs.
The Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability?: This paper proves that when alignment maps in causal abstraction are unconstrained by linearity, any neural network can be mapped to any algorithm, rendering causal abstraction trivial and uninformative. This gives rise to the "non-linear representation dilemma"—the absence of a principled trade-off between the complexity and the fidelity of alignment maps.
Sloth: Scaling Laws for LLM Skills to Predict Multi-Benchmark Performance Across Families: This paper proposes Skills Scaling Laws (Sloth), which assumes that LLM performance is driven by low-dimensional latent skills (e.g., reasoning, instruction following). By exploiting inter-benchmark correlations, Sloth constructs scaling laws that generalize across model families, enabling prediction of large-model performance on multiple benchmarks using only a small amount of family-specific data.
Base Models Know How to Reason, Thinking Models Learn When: Through unsupervised SAE clustering, this work discovers a taxonomy of reasoning mechanisms in thinking models, then activates the corresponding latent capabilities in base models via steering vectors. The resulting hybrid model recovers up to 91% of the performance gap between thinking and base models—without any weight updates—demonstrating that base models already possess reasoning capabilities, and that thinking models merely learn when to deploy them.
Causal Head Gating: A Framework for Interpreting Roles of Attention Heads in Transformers: This paper proposes Causal Head Gating (CHG), which learns a differentiable gating scalar for each attention head in a Transformer and applies positive/negative regularization to classify heads into three causal roles—facilitating, interfering, and irrelevant—without requiring manual labels or prompt templates. The framework discovers causal sub-circuits at scale and extends to Contrastive CHG for disentangling independent circuits underlying in-context learning (ICL) and instruction following.
Dense SAE Latents Are Features, Not Bugs: This paper systematically investigates frequently activating "dense latents" in sparse autoencoders (SAEs), demonstrating that they are not training artifacts but rather reflections of intrinsically dense subspaces in language model residual streams. The authors propose a six-category taxonomy of dense latents encompassing position tracking, context binding, null space, alphabetic, part-of-speech, and PCA latents.
How Do Transformers Learn Implicit Reasoning?: By training Transformers from scratch in a finely controlled symbolic environment, this work reveals that multi-hop implicit reasoning progresses through three stages: "memorization $\to$ in-distribution generalization $\to$ cross-distribution generalization." The core mechanism is not the decodability of intermediate entities, but rather their clustering consistency in the cosine space—reasoning capabilities emerge only when the representations of the same intermediate entity form tight clusters across different queries.

Browse all 81 Interpretability papers →

📦 Model Compression (146)¶

Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference: Existing KV cache eviction methods uniformly allocate budgets across all attention heads, ignoring the substantial variation in attention concentration across heads. This paper proposes Ada-KV — the first head-wise adaptive budget allocation strategy — which redistributes budget from sparse heads to dispersed heads. It provides a theoretical proof that the approach minimizes an upper bound on eviction loss, and serves as a plug-and-play improvement over existing methods across 29 datasets.
How to Build a Consistency Model: Learning Flow Maps via Self-Distillation: This paper proposes a unified self-distillation framework for directly learning flow maps (the generalized form of consistency models). By exploiting the tangent condition, any distillation scheme is converted into a direct training algorithm that requires no pretrained teacher. Three algorithm families are derived (Eulerian / Lagrangian / Progressive), among which the Lagrangian method avoids spatial gradients and bootstrapping, achieving the most stable training and best performance.
KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction: This paper proposes KVzip, a query-agnostic KV cache eviction method that quantifies the importance of each KV pair by leveraging the LLM itself to reconstruct the original context from the cached KV pairs. KVzip achieves 3–4× KV cache compression and approximately 2× reduction in FlashAttention decoding latency, while significantly outperforming existing query-aware methods in multi-query scenarios.
ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference: ChunkKV elevates the basic unit of KV cache compression from discrete tokens to semantic chunks (groups of contiguous tokens). By aggregating attention scores at the chunk level, it selects semantically intact segments for retention, and leverages the high cross-layer index similarity induced by chunking to enable layer-wise index reuse. At a 10% compression ratio, ChunkKV improves over SnapKV/PyramidKV by up to 8.7% and achieves a 26.5% throughput gain.
Twilight: Adaptive Attention Sparsity with Hierarchical Top-p Pruning: This paper proposes Twilight, which replaces fixed-budget top-k attention sparsity with a top-p (nucleus sampling) inspired approach — dynamically selecting the minimum set of tokens whose cumulative attention weights reach p%, adapting to the distribution characteristics of different attention heads. Twilight achieves up to 1.4× additional speedup over state-of-the-art sparse attention methods while maintaining accuracy.
FiRA: Can We Achieve Full-Rank Training of LLMs Under Low-Rank Constraint?: This paper proposes Fira, the first LLM training framework that achieves full-rank training (full-rank gradients + full-rank weights) under low-rank constraints. By observing that the optimizer scaling factors in low-rank and full-rank training are highly similar, Fira approximates the correction of out-of-subspace gradients using low-rank scaling factors, and employs a norm-growth limiter to prevent loss spikes. Fira outperforms LoRA and GaLore in both pretraining and fine-tuning settings.
KeyDiff: Key Similarity-Based KV Cache Eviction for Long-Context LLM Inference in Resource-Constrained Environments: This paper proposes KeyDiff — an attention-score-free KV cache eviction strategy that maintains the cache by retaining keys with the lowest average cosine similarity to other keys (i.e., geometrically most unique). Under strict memory constraints in block-wise inference settings, KeyDiff achieves ≤0.04% accuracy loss on LongBench with an 8K cache budget, while reducing end-to-end inference latency by up to 30%.
Universal Cross-Tokenizer Distillation via Approximate Likelihood Matching: This paper proposes Approximate Likelihood Matching (ALM), a principled cross-tokenizer distillation method based on binarized f-divergence, which for the first time enables effective distillation and pure distillation across fundamentally different tokenizers (e.g., subword → byte-level).
Quantization Error Propagation: Revisiting Layer-Wise Post-Training Quantization: This paper identifies a critical bottleneck in existing layer-wise PTQ methods—namely, their neglect of cross-layer accumulation and growth of quantization errors—and proposes the QEP framework, which explicitly corrects accumulated errors via error propagation and compensation, achieving substantial performance gains under extremely low-bit settings (INT2/INT3).
AI-Generated Video Detection via Perceptual Straightening: This paper proposes ReStraV, a method grounded in the perceptual straightening hypothesis—which posits that real videos form straighter trajectories in neural representation space—to detect AI-generated videos. Using temporal curvature and step-size statistics extracted from DINOv2 feature space, a lightweight classifier is trained to distinguish real from generated content, achieving 97.17% accuracy and 98.63% AUROC on VidProM with only ~48ms inference time.

Browse all 146 Model Compression papers →

🕸️ Graph Learning (54)¶

GFM-RAG: Graph Foundation Model for Retrieval Augmented Generation: This paper proposes GFM-RAG, the first graph foundation model-driven retrieval-augmented generation framework, which performs single-pass multi-hop reasoning over knowledge graphs via a query-dependent GNN. With only 8M parameters, GFM-RAG achieves zero-shot generalization to unseen datasets and substantially outperforms state-of-the-art methods on multi-hop QA retrieval benchmarks.
Heterogeneous Swarms: Jointly Optimizing Model Roles and Weights for Multi-LLM Systems: This paper proposes the Heterogeneous Swarms algorithm, which models multi-LLM systems as directed acyclic graphs (DAGs) and employs particle swarm optimization (PSO) to jointly optimize model roles (graph topology) and model weights, achieving an average improvement of 18.5% over 17 baselines across 12 tasks.
Diagnosing and Addressing Pitfalls in KG-RAG Datasets: Toward More Reliable Benchmarking: A systematic audit of 16 KGQA datasets reveals an average factual correctness of only 57% (WebQSP: 52%, MetaQA: 20%). The paper proposes KGQAGen, a framework that constructs high-quality multi-hop QA datasets via LLM-guided subgraph expansion and automatic SPARQL validation, yielding KGQAGen-10k with 96.3% accuracy. The study further demonstrates that the primary bottleneck in KG-RAG lies in retrieval rather than reasoning.
FALCON: An ML Framework for Fully Automated Layout-Constrained Analog Circuit Design: FALCON proposes an end-to-end framework for automated analog/RF circuit design via a three-stage pipeline: MLP-based topology selection, edge-centric GNN performance prediction, and differentiable layout-constrained gradient inference. Trained on a million-scale Cadence simulation dataset, the framework achieves >99% topology selection accuracy, <10% performance prediction error, and sub-second per-instance inference.
ReMindRAG: Low-Cost LLM-Guided Knowledge Graph Traversal for Efficient RAG: This paper proposes ReMindRAG, a KG-RAG system that combines LLM-guided KG traversal (node exploration + exploitation) with a training-free memory replay mechanism. The system stores LLM traversal experience in edge embeddings, enabling significant reduction in LLM calls for subsequent similar queries (~50% cost reduction) while improving answer accuracy (5%–10% gain).
Deliberation on Priors: Trustworthy Reasoning of Large Language Models on Knowledge Graphs: This paper proposes DP (Deliberation on Priors), a framework that leverages structural priors from knowledge graphs via progressive knowledge distillation to generate faithful relational paths, and validates reasoning reliability through a reasoning introspection strategy based on constraint priors, achieving new state-of-the-art performance on KGQA benchmarks.
Unifying Text Semantics and Graph Structures for Temporal Text-attributed Graphs with LLMs: This paper proposes the Cross framework, which employs LLMs to dynamically summarize the semantic evolution of node neighborhoods at strategically sampled temporal points (Temporal Reasoning Chain), then bidirectionally fuses text semantics and graph structural temporal information via a semantic-structural co-encoder. The approach achieves an average MRR improvement of 24.7% on temporal link prediction and a 3.7% AUC gain on an industrial dataset (WeChat).
Over-squashing in Spatiotemporal Graph Neural Networks: This paper provides the first formal treatment of over-squashing in spatiotemporal graph neural networks (STGNNs), uncovering a counterintuitive "temporal sink" phenomenon in causal convolutions—whereby the earliest timestep exerts the greatest influence on the final representation—and proves that time-and-space (T&S) and time-then-space (TTS) architectures are equivalent in terms of information bottlenecks, offering theoretical justification for the computationally efficient TTS design.
Practical Bayes-Optimal Membership Inference Attacks: This paper proposes BASE and G-BASE, two practical Bayes-optimal membership inference attack methods targeting i.i.d. data and graph-structured data, respectively, achieving theoretical optimality while substantially reducing computational cost.
The Underappreciated Power of Vision Models for Graph Structural Understanding: This paper reveals the severely underappreciated capability of vision models (ResNet/ViT/Swin, etc.) for graph structural understanding. By rendering graphs as images and processing them with visual encoders, these models significantly outperform GNNs in global topology perception and cross-scale generalization. The paper also introduces the GraphAbstract benchmark to systematically evaluate this finding.

Browse all 54 Graph Learning papers →

📈 Time Series (54)¶

TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning: This paper proposes TiRex, a pretrained time series forecasting model based on xLSTM. By introducing a Contiguous Patch Masking (CPM) strategy and data augmentation techniques, TiRex with only 35M parameters comprehensively outperforms larger models such as Chronos Bolt (200M) and TimesFM (500M) on the GiftEval and Chronos-ZS benchmarks, achieving state-of-the-art performance in both short- and long-horizon zero-shot forecasting.
Time-O1: Time-Series Forecasting Needs Transformed Label Alignment: This paper proposes Time-O1, which addresses the autocorrelation bias and task overload of the TMSE loss in time series forecasting by transforming label sequences into decorrelated, importance-ranked principal components. The method achieves state-of-the-art performance while remaining compatible with a wide range of forecasting models.
MIRA: Medical Time Series Foundation Model for Real-World Health Data: This paper presents MIRA, a foundation model specifically designed for irregular medical time series. Through continuous-time rotary position encoding (CT-RoPE), frequency-specific Mixture-of-Experts (MoE), and a Neural ODE-based extrapolation module, MIRA is pretrained on 454 billion observation points and achieves zero-shot forecasting performance that reduces average error by 8% and 6% in OOD and in-distribution (ID) settings, respectively.
xLSTM-Mixer: Multivariate Time Series Forecasting by Mixing via Scalar Memories: This paper proposes xLSTM-Mixer, the first architecture to combine the Extended Long Short-Term Memory network (sLSTM) with a Mixer framework. Through a three-stage design comprising temporal mixing, joint temporal-variate mixing, and multi-view mixing, the model achieves state-of-the-art performance on multivariate long-term time series forecasting while maintaining an extremely low memory footprint.
MAESTRO: Adaptive Sparse Attention and Robust Learning for Multimodal Dynamic Time Series: This paper proposes MAESTRO, a framework that addresses modality heterogeneity and arbitrary missingness in multimodal time series via symbolic tokenization, adaptive attention budgeting, sparse cross-modal attention, and dynamic MoE routing, substantially outperforming baselines under both complete and missing modality scenarios.
Structured Sparse Transition Matrices to Enable State Tracking in State-Space Models: This paper proposes PD-SSM, a structured sparse parameterization for the state transition matrix of state-space models (SSMs). The core idea is to factorize the transition matrix as a product of a column-wise one-hot matrix P and a complex diagonal matrix D (i.e., $A = PD$), achieving expressiveness equivalent to unstructured (dense) SSMs while retaining computational efficiency comparable to diagonal SSMs at $\Theta(LN)$. A single layer suffices to simulate any $N$-state finite-state automaton (FSA). The paper provides theoretical guarantees on BIBO stability and optimal state dimensionality, with strong empirical results on FSA simulation, multivariate time-series classification, long-sequence benchmarks, and natural-language state-tracking tasks.
Learning with Calibration: Exploring Test-Time Computing of Spatio-Temporal Forecasting: This paper proposes ST-TTC, a lightweight test-time computing paradigm that corrects periodic biases in spatio-temporal forecasting during inference via a frequency-domain phase-amplitude calibrator and a flash gradient update mechanism, consistently improving the performance of diverse backbone models without modifying their architectures.
Multi-Scale Finetuning for Encoder-based Time Series Foundation Models: This paper proposes MSFT (Multi-Scale FineTuning), which leverages causal analysis to reveal that naive fine-tuning suffers from scale confounding, and designs a multi-scale modeling framework for efficient fine-tuning of encoder-based time series foundation models, significantly outperforming both naive fine-tuning and from-scratch SOTA methods.
Improving Time Series Forecasting via Instance-aware Post-hoc Revision (PIR): PIR proposes an instance-aware post-hoc revision framework that identifies poorly predicted instances via uncertainty estimation and applies a residual combination of local correction (covariate + exogenous variable Transformer) and global correction (retrieval-based weighted average over similar training instances) as a plug-and-play module, reducing SparseTSF MSE by 25.87% and PatchTST MSE by 8.99%.
OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales: OmniCast is proposed as a weather forecasting method that combines a masked generative framework with a latent diffusion model. By jointly generating future weather sequences rather than iterating autoregressively, it mitigates error accumulation, achieves state-of-the-art performance at the subseasonal-to-seasonal (S2S) scale, remains competitive for medium-range forecasting, and offers inference speeds 10–20× faster.

Browse all 54 Time Series papers →

🏥 Medical Imaging (78)¶

DyG-Mamba: Continuous State Space Modeling on Dynamic Graphs: DyG-Mamba introduces continuous state space models (SSMs) into dynamic graph learning. It proposes a temporal span-aware continuous SSM that models irregular time intervals via an exponential decay function inspired by the Ebbinghaus forgetting curve, combined with input-dependent parameters constrained by spectral norm for Lipschitz robustness. The method achieves an average rank of 2.42 across 12 dynamic graph benchmarks (vs. DyGFormer's 2.92) while maintaining $O(bdL)$ linear complexity.
BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals: This paper proposes BrainOmni—the first brain signal foundation model that unifies EEG and MEG—by discretizing heterogeneous brain signals into a unified token space via BrainTokenizer (incorporating a physical Sensor Encoder), followed by self-supervised masked prediction pretraining with a Criss-Cross Transformer. The model achieves an 11.7 percentage-point improvement on Alzheimer's disease detection and demonstrates zero-shot reconstruction generalization to completely unseen devices.
3D-RAD: A Comprehensive 3D Radiology Med-VQA Dataset with Multi-Temporal Analysis and Diverse Diagnostic Tasks: This paper introduces 3D-RAD — the first large-scale 3D medical VQA benchmark, comprising 170K CT-based question-answer pairs across six clinical task categories (including a novel multi-temporal diagnosis task), accompanied by a 136K training set. The benchmark reveals critical deficiencies of existing VLMs in 3D temporal reasoning.
Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens: The first multimodal brain foundation model that unifies structural morphology (T1 sMRI) and functional dynamics (fMRI), compressing high-dimensional neuroimaging data into compact 1D token representations via Geometric Harmonics Pre-alignment and Temporally Adaptive Patch Embedding (TAPE). The model consistently outperforms prior methods on neurodevelopmental/neurodegenerative disease diagnosis and cognitive prediction tasks.
A Unified Solution to Video Fusion: From Multi-Frame Learning to Benchmarking: This paper proposes UniVF, the first unified video fusion framework based on multi-frame learning, optical flow feature warping, and temporal consistency loss, along with VF-Bench, the first video fusion benchmark covering four major fusion tasks (multi-exposure, multi-focus, infrared-visible, and medical), achieving state-of-the-art performance across all sub-tasks.
ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression: By proposing a systematic feature suppression framework—rather than cue-conflict experiments—this work re-evaluates the feature reliance of CNNs, finding that CNNs are not inherently texture-biased but instead rely primarily on local shape features; moreover, feature reliance patterns differ substantially across domains (CV/MI/RS).
EndoBench: A Comprehensive Evaluation of Multi-Modal Large Language Models for Endoscopy Analysis: This paper introduces EndoBench, the first comprehensive MLLM evaluation benchmark covering 4 endoscopic scenarios, 12 clinical tasks, and 5 levels of visual prompt granularity, comprising 6,832 clinically validated VQA pairs. Evaluation of 23 MLLMs reveals that commercial models generally outperform open-source and medical-specific counterparts, yet all remain below human expert performance.
PhysioWave: A Multi-Scale Wavelet-Transformer for Physiological Signal Representation: This paper proposes PhysioWave, a multi-scale Transformer architecture based on learnable wavelet decomposition and frequency-guided masking. It establishes, for the first time, large-scale pretrained foundation models for EMG and ECG, and achieves state-of-the-art performance on both unimodal and multimodal physiological signal tasks through a multimodal fusion framework.
Generalizable, Real-Time Neural Decoding with Hybrid State-Space Models: POSSM proposes a hybrid SSM-attention architecture that combines spike-level tokenization with a recurrent state-space model backbone, achieving generalizable real-time neural decoding with inference speeds up to 9× faster than Transformers while maintaining comparable accuracy.
MoRE-Brain: Routed Mixture of Experts for Interpretable and Generalizable Cross-Subject fMRI Visual Decoding: This paper proposes MoRE-Brain, a neuroscience-inspired fMRI visual decoding framework that employs a hierarchical Mixture-of-Experts (MoE) architecture to simulate the specialized processing of the brain's visual pathway. Combined with a dynamic temporal-spatial dual-routing mechanism that guides image generation via a diffusion model, MoRE-Brain achieves high-fidelity reconstruction while enabling efficient cross-subject generalization and unprecedented mechanistic interpretability.

Browse all 78 Medical Imaging papers →

🩺 Medical LLM (17)¶

PatientSim: A Persona-Driven Simulator for Realistic Doctor-Patient Interactions: This paper introduces PatientSim — an LLM-based patient simulator grounded in real MIMIC clinical data and a four-dimensional persona framework (personality, language proficiency, medical history recall, and cognitive confusion), generating 37 unique persona combinations. The system is evaluated across 8 LLMs for factual accuracy and persona fidelity, and validated by 4 clinical experts with a mean quality score of 3.89/4.
Time-IMM: A Dataset and Benchmark for Irregular Multimodal Multivariate Time Series: This work constructs Time-IMM — the first multimodal multivariate time series benchmark that categorizes irregularity according to causal mechanisms (9 irregularity types organized into three classes: Trigger, Constraint, and Artifact, spanning 9 datasets). An accompanying forecasting library, IMM-TSF, supports asynchronous multimodal fusion. Experiments demonstrate that explicitly modeling multimodal information reduces MSE by 6.71% on average across irregular time series settings, with a maximum improvement of 38.38%.
Document Summarization with Conformal Importance Guarantees: This work presents the first application of Conformal Prediction to document summarization. By calibrating a threshold on sentence importance scores, it provides rigorous statistical guarantees on user-controllable coverage ($1-\alpha$) and recall ($\beta$) for extractive summaries. The method is model-agnostic and requires only a small calibration set.
HealthSLM-Bench: Benchmarking Small Language Models for Mobile and Wearable Healthcare Monitoring: The first benchmark systematically evaluating small language models (SLMs, 1–4B parameters) on mobile and wearable health monitoring tasks, covering zero-shot, few-shot, and instruction fine-tuning paradigms, with on-device deployment validated on an iPhone.
CGBench: Benchmarking Language Model Scientific Reasoning for Clinical Genetics Research: This paper introduces CGBench, a clinical genetics benchmark grounded in ClinGen expert annotations, designed to evaluate the scientific literature reasoning capabilities of LLMs from both variant and gene curation perspectives. The benchmark encompasses three tasks—evidence scoring, evidence verification, and experimental evidence extraction—and finds that reasoning models perform best on fine-grained tasks but underperform non-reasoning models on high-level judgments.
Shallow Robustness, Deep Vulnerabilities: Multi-Turn Evaluation of Medical LLMs: This paper proposes the MedQA-Followup framework to systematically evaluate the multi-turn robustness of medical LLMs. It reveals that models exhibit acceptable performance under single-turn perturbations (shallow robustness), yet accuracy can catastrophically drop from 91.2% to 13.5% under multi-turn follow-up challenges (deep vulnerability). Notably, indirect contextual manipulation proves more destructive than direct incorrect suggestions.
AI Should Sense Better, Not Just Scale Bigger: Adaptive Sensing as a Paradigm Shift: Inspired by biological sensory systems, this position paper argues that AI research must shift from simply scaling models to optimizing inputs—by dynamically adjusting sensor-level parameters (exposure, gain, multimodal configuration, etc.) to produce inputs most favorable to the model. Under ideal sensor adaptation, a small model (EfficientNet-B0, 5M parameters) can outperform a large model (OpenCLIP-H, 632M parameters), and the paper proposes a progressive formalization framework ranging from single-shot perception to closed-loop perception–action coupling.
CureAgent: A Training-Free Executor-Analyst Framework for Clinical Reasoning: CureAgent proposes an Executor-Analyst collaborative framework that decouples precise tool invocation (TxAgent/Llama-8B as Executor) from high-level clinical reasoning (Gemini 2.5 as Analyst). Combined with a Stratified Ensemble Late Fusion topology that preserves evidence diversity, the system achieves 83.8% accuracy on CURE-Bench without end-to-end fine-tuning, and reveals two critical scaling findings: the context–performance paradox and the curse of dimensionality in action space.
Demo: Guide-RAG: Evidence-Driven Corpus Curation for Retrieval-Augmented Generation in Long COVID: This paper systematically evaluates six RAG corpus configurations for Long COVID clinical QA. The GS-4 configuration—combining clinical guidelines with high-quality systematic reviews—consistently outperforms both single-guideline and large-scale literature retrieval baselines across faithfulness, relevance, and comprehensiveness. The authors further introduce the Guide-RAG framework and the LongCOVID-CQ evaluation dataset.
Faithful Summarization of Consumer Health Queries: A Cross-Lingual Framework with LLMs: A framework combining TextRank-based extractive sentence selection and medical Named Entity Recognition (NER) is proposed to guide LLMs in generating faithful medical summaries. By fine-tuning LLaMA-2-7B on the English MeQSum and Bengali BanglaCHQ-Summ datasets, consistent improvements in both quality and faithfulness are achieved, with SummaC reaching 0.57 and human evaluation showing that 82% of the summaries retain key medical information.

Browse all 17 Medical LLM papers →

🧬 Computational Biology (77)¶

Remasking Discrete Diffusion Models with Inference-Time Scaling: This paper proposes the ReMDM sampler, which enables iterative error correction in discrete mask diffusion models by allowing already-decoded tokens to be remasked during generation. This mechanism supports inference-time compute scaling and yields substantial quality improvements on text, image, and molecular design tasks.
KLASS: KL-Guided Fast Inference in Masked Diffusion Models: This paper proposes KLASS (KL-Adaptive Stability Sampling), a training-free sampling method that leverages token-level KL divergence and confidence scores to identify stable tokens for parallel decoding, achieving up to 2.78× speedup on masked diffusion models without sacrificing—and in many cases improving—generation quality.
Consistent Sampling and Simulation: Molecular Dynamics with Energy-Based Diffusion Models: This paper identifies an inconsistency between sampling and simulation in diffusion models (particularly at small diffusion timesteps), proposes a Fokker-Planck-based regularization term to enforce consistency, and combines it with a time-partitioned Mixture-of-Experts (MoE) strategy to achieve consistent and efficient sampling and molecular dynamics simulation across multiple biomolecular systems.
Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling: This paper proposes Energy Matching, which unifies flow matching and energy-based models via a single time-independent scalar potential field: far from the data manifold, the model performs efficient transport along optimal transport paths; near the manifold, it transitions to a Boltzmann equilibrium distribution for likelihood modeling. The method achieves FID 3.34 on CIFAR-10, substantially outperforming existing EBMs by more than 50%.
Mol-LLaMA: Towards General Understanding of Molecules in Large Molecular Language Models: This paper proposes Mol-LLaMA, a large molecular language model for general molecular understanding. By designing three types of instruction data and a 2D-3D molecular representation fusion module, Mol-LLaMA surpasses GPT-4o in molecular feature understanding while exhibiting interpretability and reasoning capabilities.
Curly Flow Matching for Learning Non-gradient Field Dynamics: The authors propose Curly Flow Matching (Curly-FM), which designs a Schrödinger Bridge problem with a non-zero reference drift. This allows flow matching to learn non-gradient field dynamics, such as periodic and rotational behaviors, breaking the limitation of traditional methods that can only model gradient fields.
Split Gibbs Discrete Diffusion Posterior Sampling: This paper proposes SGDD (Split Gibbs Discrete Diffusion), a plug-and-play posterior sampling algorithm for discrete diffusion models based on the split Gibbs sampling principle. By introducing auxiliary variables and a Hamming-distance-based regularization potential, SGDD decomposes posterior sampling into alternating likelihood and prior sampling steps, achieving substantial improvements over baselines on DNA sequence design, discrete image inverse problems, and music infilling tasks.
Amortized Sampling with Transferable Normalizing Flows: This work proposes Prose — a 285M-parameter all-atom transferable normalizing flow based on the TarFlow architecture, trained on 21,700 short-peptide MD trajectories (totaling 4.3 ms of simulation time). Prose enables zero-shot uncorrelated proposal sampling for arbitrary short-peptide systems, outperforms MD baselines under equal energy evaluation budgets, and generates samples 4,000× faster than the prior transferable Boltzmann generator (TBG).
Constrained Discrete Diffusion: This paper proposes CDD (Constrained Discrete Diffusion), which embeds a differentiable constrained optimization projection operator into the denoising process of discrete diffusion models. Without retraining, CDD enforces sequence-level constraints at sampling time, achieving zero constraint violations across three task categories: toxic text generation, molecular design, and instruction following.
Flow Density Control: Generative Optimization Beyond Entropy-Regularized Fine-Tuning: This paper proposes Flow Density Control (FDC), which generalizes the fine-tuning of pretrained flow/diffusion models from KL-regularized expected reward maximization to a unified framework supporting arbitrary distributional utility functions with arbitrary divergence regularization. The approach decomposes nonlinear objectives into a sequence of linear fine-tuning subproblems and provides convergence guarantees.

Browse all 77 Computational Biology papers →

⚛️ Physics & Scientific Computing (64)¶

Physics-Constrained Flow Matching: Sampling Generative Models with Hard Constraints: This paper proposes Physics-Constrained Flow Matching (PCFM), a zero-shot inference framework that enforces arbitrary nonlinear equality constraints to machine precision during sampling from pretrained flow matching models. The framework alternates among forward shooting with projection, OT-interpolation backward updates, and relaxed penalty correction at each sub-step, achieving up to 99.5% improvement over baselines on PDE problems involving shocks and discontinuities.
Guided Diffusion Sampling on Function Spaces with Applications to PDEs: This paper proposes FunDPS (Function-space Diffusion Posterior Sampling), which trains an unconditional diffusion model in function space and performs plug-and-play posterior sampling for PDE inverse problems via gradient guidance at inference time. Theoretically, it extends the Tweedie formula to infinite-dimensional Banach spaces. Empirically, across 5 PDE tasks with only 3% observations, FunDPS achieves 32% higher accuracy on average than DiffusionPDE while reducing the number of sampling steps by 4×.
FlashMD: Long-Stride, Universal Prediction of Molecular Dynamics: FlashMD is proposed as a GNN-based framework that directly predicts the positional and momentum evolution of molecular dynamics trajectories with long strides, achieving time steps 1–2 orders of magnitude larger than those of conventional MD integrators. The architecture incorporates Hamiltonian dynamics constraints and generalizes to arbitrary thermodynamic ensembles and universal chemical systems.
Tropical Attention: Neural Algorithmic Reasoning for Combinatorial Algorithms: Tropical Attention replaces softmax dot-product attention with tropical algebraic geometry, performing piecewise-linear reasoning in tropical projective space to align with the polyhedral decision structures of combinatorial algorithms. It is the first approach to extend neural algorithmic reasoning to NP-hard problems, comprehensively outperforming softmax baselines across three OOD generalization axes: length, magnitude, and noise.
High-order Equivariant Flow Matching for Density Functional Theory Hamiltonian Prediction: This paper proposes QHFlow, the first method to apply conditional flow matching to density functional theory (DFT) Hamiltonian matrix prediction. By designing high-order SE(3)-equivariant vector fields and symmetry-aware prior distributions, QHFlow reduces Hamiltonian prediction error by 73% on MD17 and accelerates DFT computation by 54% when used as an SCF initializer.
GyroSwin: 5D Surrogates for Gyrokinetic Plasma Turbulence Simulations: This work presents GyroSwin, the first scalable 5D neural surrogate model for gyrokinetic plasma turbulence. It extends the Swin Transformer to the 5D gyrokinetic phase space, employs cross-attention for 3D↔5D interaction, and adopts channelwise mode separation to capture zonal flows. GyroSwin achieves higher accuracy than conventional quasilinear methods while being three orders of magnitude faster than the numerical solver GKW.
A Regularized Newton Method for Nonconvex Optimization with Global and Local Complexity Guarantees: This paper proposes a novel class of regularizers constructed from current and historical gradients, combined with a conjugate gradient method equipped with negative-curvature detection to solve the regularized Newton equation. Within an adaptive framework that requires no prior knowledge of the Hessian Lipschitz constant, the method simultaneously achieves, for the first time, the optimal global iteration complexity of $O(\epsilon^{-3/2})$ and a quadratic local convergence rate.
Adaptive Stochastic Coefficients for Accelerating Diffusion Sampling: By theoretically analyzing the complementary weaknesses of ODE and SDE solvers (ODE solvers accumulate irreducible gradient errors; SDE solvers amplify discretization errors at large step sizes), this paper proposes AdaSDE—a method that introduces a learnable stochastic coefficient $\gamma_i$ at each denoising step to control noise injection intensity. Optimized via lightweight distillation, AdaSDE achieves state-of-the-art FID of 4.18 on CIFAR-10 and 8.05 on FFHQ at 5 NFE.
EddyFormer: Accelerated Neural Simulations of Three-Dimensional Turbulence at Scale: EddyFormer is a Transformer architecture based on the Spectral Element Method (SEM) that decomposes the flow field into two parallel streams — LES (large-scale) and SGS (small-scale) — achieving DNS-level accuracy on 3D turbulence at $256^3$ resolution with a 30× speedup, while generalizing well to unseen domains 4× larger.
Neural Emulator Superiority: When Machine Learning for PDEs Surpasses its Training Data: This work challenges the prevailing assumption that the accuracy of neural PDE emulators is bounded by that of their training data (i.e., the numerical solver). It discovers and rigorously defines the phenomenon of emulator superiority—neural networks trained solely on low-accuracy solver data can, when evaluated against high-accuracy reference solutions, outperform the very solver that generated their training data.

Browse all 64 Physics & Scientific Computing papers →

🌍 Earth Science (6)¶

ControlFusion: A Controllable Image Fusion Framework with Language-Vision Degradation Prompts: This paper proposes ControlFusion, a controllable infrared-visible image fusion framework based on language-vision degradation prompts. It employs a physics-driven degradation imaging model to simulate compound degradations, and uses a prompt-modulated network to perform dynamic restoration and fusion, achieving comprehensive state-of-the-art performance under both real-world and compound degradation scenarios.
A Probabilistic U-Net Approach to Downscaling Climate Simulations: This work presents the first application of a probabilistic U-Net to statistical climate downscaling (16× super-resolution). By sampling from a variational latent space, the model generates ensemble forecasts for uncertainty quantification. The paper systematically compares four training objectives — WMSE, MS-SSIM, WMSE-MS-SSIM, and afCRPS — revealing complementary trade-offs between extreme event capture and fine-scale spatial variability preservation.
Adaptive Online Emulation for Accelerating Complex Physical Simulations: This paper proposes Adaptive Online Emulation (AOE), a framework that dynamically trains an ELM-based neural network surrogate model during physical simulation execution to replace expensive computational components, requiring no offline pretraining. On an exoplanetary atmospheric simulation, AOE achieves an 11.1× speedup (91% time savings) with only ~0.01% accuracy loss.
Power Ensemble Aggregation for Improved Extreme Event AI Prediction: This paper proposes an adaptive ensemble aggregation method based on the power mean. By applying nonlinear aggregation (power exponent $p>1$) to the score of ensemble members from generative weather prediction models, the method significantly improves classification performance for extreme high-temperature events, with greater gains at higher quantile thresholds.
Predicting Public Health Impacts of Electricity Usage: This paper proposes HealthPredictor, an AI pipeline that maps electricity consumption end-to-end to public health damages (measured in $/MWh), comprising three modules: fuel mix prediction, air quality conversion, and health impact assessment. Health-driven optimization significantly reduces health impact prediction error compared to fuel-mix-driven baselines, and achieves a 24–42% reduction in health damages in an EV charging scheduling case study.
Reasoning With a Star: A Heliophysics Dataset and Benchmark for Agentic Scientific Reasoning: This paper introduces Reasoning With a Star (RWS), a 158-question scientific reasoning benchmark derived from NASA heliophysics summer school problem sets, spanning three answer types (numeric/symbolic/textual). Paired with a unit-aware grader, it evaluates four multi-agent coordination paradigms (HMAW/PACE/PHASE/SCHEMA) and finds that no single paradigm dominates across all tasks — the systems-engineering-inspired SCHEMA achieves the strongest performance on tasks requiring rigorous constraint validation.

📡 Signal & Communications (5)¶

Angular Steering: Behavior Control via Rotation in Activation Space: This paper proposes Angular Steering, which unifies LLM activation steering as rotation operations within a fixed 2D subspace — providing a continuous, fine-grained, norm-preserving behavior control knob spanning 0°–360° via rotation angle. The framework subsumes activation addition and directional ablation as special cases of rotation, and demonstrates robust behavior control on Llama 3 / Qwen 2.5 / Gemma 2 (3B–14B).
Contrastive Consolidation of Top-Down Modulations Achieves Sparsely Supervised Continual Learning: This paper proposes Task-Modulated Contrastive Learning (TMCL), inspired by top-down modulations in the neocortex. TMCL integrates sparse label information (as few as 1% labels) via affine modulation during continual learning, then consolidates the modulation information into feedforward weights through contrastive learning, surpassing both unsupervised and supervised baselines on class-incremental and transfer learning benchmarks.
Feature-aware Modulation for Learning from Temporal Tabular Data: This paper argues that the core challenge in temporal tabular learning is not simply "adding a time embedding," but rather that the semantics of many features drift over time. To address this, the paper proposes feature-aware modulation, which uses temporal context to dynamically generate per-feature shift, scale, and nonlinear shape parameters, re-aligning cross-temporal semantics. The approach enables deep models to consistently outperform GBDT on average rank for the first time on the TabReD benchmark.
Memory-Integrated Reconfigurable Adapters: A Unified Framework for Settings with Multiple Tasks: MIRA embeds Hopfield-style associative memory modules into each layer of a ViT, storing and retrieving LoRA adapter weights as key-value pairs. Through a two-stage training procedure (Adaptation + Consolidation), a single unified architecture simultaneously addresses Domain Generalization (DG), Class-Incremental Learning (CIL), and Domain-Incremental Learning (DIL), achieving substantial improvements over task-specific methods across multiple benchmarks.
Masked Symbol Modeling for Demodulation of Oversampled Baseband Communication Signals: This paper proposes Masked Symbol Modeling (MSM), transplanting BERT's masked prediction paradigm to the communication physical layer. It reframes inter-symbol contributions from pulse shaping as "contextual information," training a Transformer on clean oversampled baseband signals to learn waveform structure, and leveraging the learned context at inference time to recover symbols corrupted by impulsive noise.

👥 Social Computing (20)¶

OS-Harm: A Benchmark for Measuring Safety of Computer Use Agents: This paper presents OS-Harm, the first safety benchmark targeting general-purpose computer use agents (beyond browser-only settings), covering 150 tasks across three risk categories — deliberate user misuse, prompt injection attacks, and model misbehavior. Evaluations reveal that frontier models (o4-mini, Claude 3.7 Sonnet, Gemini 2.5 Pro, etc.) broadly comply with harmful instructions (up to 70% unsafe rate) and exhibit a 20% compliance rate against basic prompt injection attacks.
Worse than Zero-shot? A Fact-Checking Dataset for Evaluating the Robustness of RAG Against Misleading Retrievals: This paper introduces RAGuard, the first benchmark dataset to systematically evaluate the robustness of RAG systems against misleading retrieved content. By constructing a realistic retrieval corpus from Reddit — containing supporting, misleading, and unrelated documents — it demonstrates that all tested LLM-RAG systems perform worse than a zero-shot baseline when exposed to misleading retrievals, whereas human annotators maintain consistent judgment.
AVerImaTeC: A Dataset for Automatic Verification of Image-Text Claims with Evidence from the Web: AVerImaTeC introduces the first image-text fact-checking dataset with complete evidence annotation — 1,297 real-world image-text claims, a 5-stage annotation pipeline (extraction → QA reasoning → sufficiency check → iterative refinement → second check), and temporally constrained evidence (to prevent temporal leakage). The baseline system achieves 82% accuracy with ground-truth evidence, but drops to 15–25% under automatic evidence retrieval, revealing the substantial challenges of image-text verification.
Any Large Language Model Can Be a Reliable Judge: Debiasing with a Reasoning-based Bias Detector: This paper proposes the Reasoning-based Bias Detector (RBD), a plug-and-play debiasing module for LLM judges. By externally detecting four types of evaluation bias (verbosity, position, bandwagon, and sentiment), RBD generates structured feedback with reasoning chains to guide judges toward self-correction. RBD-8B achieves an average accuracy improvement of 18.5% and consistency improvement of 10.9% across 8 LLM judges.
GraphKeeper: Graph Domain-Incremental Learning via Knowledge Disentanglement and Preservation: GraphKeeper is proposed to address catastrophic forgetting in Graph Domain-Incremental Learning (Graph Domain-IL) through three components: domain-specific LoRA parameter isolation, intra/inter-domain disentanglement, and ridge regression-based deviation-free knowledge preservation. It outperforms the second-best method by 6.5%–16.6% and can be seamlessly integrated with graph foundation models.
IF-GUIDE: Influence Function-Guided Detoxification of LLMs: This paper proposes IF-Guide, which leverages influence functions to identify toxic content in training data at the token granularity and actively suppresses the model from learning toxic behaviors during pre-training or fine-tuning via a penalty-based training objective, substantially outperforming passive alignment methods such as DPO and RAD.
Precise Information Control in Long-Form Text Generation: This paper proposes the Precise Information Control (PIC) task, which requires LLMs to generate long-form text that strictly adheres to a given set of claims (neither omitting nor adding information). The authors construct PIC-Bench to evaluate 8 tasks, finding that over 70% of outputs from state-of-the-art models contain faithfulness hallucinations. Through weakly supervised preference data construction combined with DPO training, the proposed PIC-LM improves the F1 of an 8B model from 69.1% to 91.0%.
DATE-LM: Benchmarking Data Attribution Evaluation for Large Language Models: DATE-LM introduces the first unified benchmark for evaluating data attribution methods in LLMs. Through three application-driven tasks—training data selection, toxicity filtering, and factual attribution—it systematically compares multiple attribution approaches, finding that no single method dominates across all tasks and that simple baselines can match attribution methods in certain settings.
VDRP: Visual Diversity and Region-aware Prompt Learning for Zero-shot HOI Detection: This paper proposes the VDRP framework, which addresses two core challenges in zero-shot HOI detection — intra-class visual diversity and inter-class visual entanglement — through visual diversity-aware prompt learning (via group-level variance injection and Gaussian perturbation) and region-aware prompt augmentation (via LLM-generated regional concept retrieval).
Noise-Robustness Through Noise: A Framework Combining Asymmetric LoRA with Poisoning MoE: This paper proposes LoPE, which designates a dedicated "poisoning expert" within an asymmetric LoRA architecture to absorb injected noise during training; at inference time, this expert is masked so that only the clean experts contribute to the output — achieving noise robustness through noise itself, entirely without data cleaning.

Browse all 20 Social Computing papers →

📂 Others (123)¶

FSNet: Feasibility-Seeking Neural Network for Constrained Optimization with Guarantees: This paper proposes FSNet, a framework that integrates differentiable feasibility-seeking steps into neural networks. By minimizing constraint violations via unconstrained optimization, FSNet guarantees constraint satisfaction while supporting end-to-end training. It significantly outperforms traditional solvers in speed across convex/non-convex and smooth/non-smooth problems while maintaining feasibility.
Continuous Thought Machines: This paper proposes the Continuous Thought Machine (CTM), which generates neuron-level temporal dynamics via privately parameterized Neuron-Level Models (NLMs) and employs a neural synchrony matrix as the core latent representation. The model demonstrates complex reasoning, adaptive computation, and interpretable attention behavior on tasks including maze solving, ImageNet classification, and parity checking.
Adjoint Schrödinger Bridge Sampler: This paper proposes the Adjoint Schrödinger Bridge Sampler (ASBS), which reinterprets the Schrödinger Bridge problem as a stochastic optimal control (SOC) problem. This eliminates the memoryless condition required by prior diffusion samplers, supports arbitrary source distributions (e.g., Gaussian, harmonic priors), and employs a scalable matching objective without importance weight estimation. ASBS consistently outperforms prior methods on multi-particle energy functions and molecular conformation generation.
Generalized Linear Mode Connectivity for Transformers: This paper proposes a unified symmetry framework (a four-level hierarchy of permutation, semi-permutation, orthogonal, and invertible transformations) to achieve zero or near-zero barrier linear mode connectivity (LMC) on Vision Transformers and GPT-2 for the first time, and further extends the framework to multi-model merging and heterogeneous-width alignment.
Modeling Cell Dynamics and Interactions with Unbalanced Mean Field Schrödinger Bridge: This paper proposes the Unbalanced Mean Field Schrödinger Bridge (UMFSB) framework and the CytoBridge deep learning algorithm, which simultaneously model unbalanced stochastic cell dynamics and cell-cell interactions from sparse temporal snapshot data.
Depth-Bounds for Neural Networks via the Braid Arrangement: This paper proves that, under $\mathcal{B}_d^0$-conforming constraints, exactly representing $\max\{0, x_1, \ldots, x_d\}$ with a ReLU network requires $\Omega(\log \log d)$ layers—the first non-constant depth lower bound without weight restrictions. It also shows that rank-(3,2) maxout networks can compute the maximum of 7 values, demonstrating that the standard upper bound is not tight.
A Theoretical Framework for Grokking: Interpolation followed by Riemannian Norm Minimisation: This paper rigorously proves the mechanism behind grokking from a purely optimization-theoretic perspective. Gradient flow with small weight decay exhibits two-phase dynamics in the $\lambda\to 0$ limit: rapid convergence to the critical manifold $\mathcal{M}$ of the training loss, followed by a Riemannian gradient flow along the manifold minimizing the $\ell_2$ norm at timescale $t\approx 1/\lambda$, thereby inducing delayed generalization.
Learning (Approximately) Equivariant Networks via Constrained Optimization: This paper proposes ACE (Adaptive Constrained Equivariance), a framework that formulates equivariant neural network training as a constrained optimization problem. Via dual methods, ACE automatically and progressively transitions from a flexible non-equivariant model to an equivariant one, adapting to both fully and partially symmetric data without manual hyperparameter tuning.
Uncertainty Estimation by Flexible Evidential Deep Learning: This paper proposes $\mathcal{F}$-EDL, which generalizes the Dirichlet distribution in EDL to a Flexible Dirichlet (FD) distribution for modeling class probability distributions. This approach significantly enhances the generalization of uncertainty estimation under complex scenarios such as noise, long-tail distributions, and distribution shift, while preserving the efficiency of a single forward pass.
Dense Associative Memory with Epanechnikov Energy: This paper proposes a log-sum-ReLU (LSR) energy function based on the Epanechnikov kernel as a replacement for the conventional log-sum-exp (LSE) energy in Dense Associative Memory. For the first time, it achieves the coexistence of exact retrieval of all stored patterns and the emergence of novel creative local minima, while preserving exponential memory capacity.

Browse all 123 Others papers →