Skip to content

🧠 NeurIPS2026 Accepted Papers

4078 NeurIPS2026 paper notes covering Optimization & Theory (313), LLM Agent (270), Reinforcement Learning (237), Safety (208), Image Generation (206), Model Compression (205), Learning Theory (202), Multimodal VLM (175) and other 54 areas. Each note has TL;DR, motivation, method, experiments, highlights, and limitations — 5-minute reads of core ideas.


💡 LLM Reasoning (111)

A Mechanistic Analysis of Looped Reasoning Language Models

Through a rigorous mechanistic analysis of latent trajectories and attention dynamics, this paper reveals that recurrent Transformer layers converge to distinct cyclic fixed points, stabilizing attention patterns across recurrences and periodically replicating feedforward stages of inference within each recurrent block.

A Theoretical Framework for Self-Play Theorem Proving Algorithms

This paper introduces the first theoretical framework for LLM-based self-play theorem proving, proving that a prover-conjecturer system over a well-connected theorem graph achieves exponential growth in provable theorems, and designing a diffusion similarity-guided conjecturer via contrastive learning to resolve theorem clustering collapse.

A Theory of Online Learning with Autoregressive Chain-of-Thought Reasoning

This paper develops the theoretical foundations of online learning with autoregressive chain-of-thought reasoning, establishing an exact taxonomy of end-to-end mistake bound growth from constant to logarithmic \(\Theta(L(F)\log M)\) in generation horizon \(M\), while proving that full chain-of-thought feedback completely eliminates any dependence on \(M\).

Adaptive Generate-Rank-Verify: Inference-Time Search with Costly Verification

Addressing the compute-allocation dilemma between cheap reward proxies and expensive gold verifiers in LLM reasoning, this paper establishes a generative active search theoretical framework and introduces ADAP, an online shellwise adaptive policy that eliminates the need for offline difficulty calibration while matching 100% success at up to 5.5× lower cost with constant-factor optimality guarantees.

Advancing Entropy-Level Credit Assignment in RLVR via Proximal Entropy Policy Optimization

PEPO weights token updates in successful reasoning trajectories by uncertainty relative to their neighborhoods while preserving each trajectory's total weight, outperforming the compared GRPO and global-entropy baselines in the main mathematical reasoning experiments, with explicit limits on trend robustness and cross-task generalization.

An Enigma of Artificial Reason: Investigating the Production-Evaluation Gap in Large Reasoning Models

This paper uncovers a severe "production-evaluation gap" in Large Reasoning Models (LRMs): while achieving near-perfect accuracy when producing solutions from scratch, frontier models score as low as 48% when evaluating solutions with trivial reasoning flaws but valid answers (VAIR), fundamentally driven by an answer confirmation bias that overrides internal validity representations.

Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information

This paper demonstrates that privileged on-policy self-distillation inherently amounts to conditional pointwise mutual information (PMI), which systematically inflates shortcut tokens and penalizes deliberation; to resolve this, the authors propose Anti-Self-Distillation (AntiSD) by ascending Jensen-Shannon divergence under an entropy-triggered gate, achieving 2-10x training speedups and up to +11.5 pp accuracy gains over GRPO without external models.

Are Your Reasoning Models Reasoning or Guessing? A Mechanistic Analysis of Hierarchical Reasoning Models

Through a mechanistic analysis of latent-space trajectories in Hierarchical Reasoning Models (HRM), this paper reveals that HRM operates by "guessing" fixed points rather than continuous reasoning, and introduces data mixing, input perturbations, and model bootstrapping to elevate Sudoku-Extreme accuracy from 55.0% to 96.9%.

Ares: Adaptive Reasoning Effort Selection for Efficient LLM Agents

Ares introduces a lightweight per-step reasoning router that adaptively selects the minimum sufficient reasoning effort level for each step in multi-turn agent interactions, cutting reasoning tokens by up to 52.7% while preserving task success rates.

Asking the Right Questions: Improving Reasoning with Generated Stepping Stones

The paper introduces ARQ (Asking the Right Questions), a framework that decouples question generation from problem solving by producing intermediate stepping stones, and establishes stepping stone formulation as an optimizable post-training task via Monte Carlo expected reward estimation, SFT, and DPO.

Browse all 111 LLM Reasoning papers →


🦾 LLM Agent (270)

3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code

3DCodeBench introduces a comprehensive benchmark spanning 212 semantic categories and 26K multimodal prompt-code-mesh triplets paired with the 3DCodeArena human preference platform, systematically dissecting the execution capabilities, API adaptabilities, and physical plausibility bottlenecks of 12 frontier VLMs.

A Few Teacher Steps Go a Long Way: Cost-Efficient On-Policy Data Augmentation for Agent Post-Training

This paper formulates on-policy data augmentation for LLM agents as a budget-allocation problem, proving that bounded few-step teacher continuations at learner-induced states outperform pure behavioral cloning and heavily filtered baselines at matched supervision costs.

A Matter of TASTE: Improving Coverage and Difficulty of Agent Benchmarks

Addressing the rapid saturation and narrow procedural coverage of existing conversational agent benchmarks, this paper reverses the task construction process via TASTE, which explores tool-sequence space using an adaptive contrastive n-gram model, selects representative sequences via weighted clustering, and applies adversarial task evolution to build the challenging, high-coverage \(\tau^c\)-Bench.

Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security

Presents Security Arena, a 21-scenario benchmark pairing autonomous LLM attackers with fresh-session LLM defenders, demonstrating that single-round evaluation severely underestimates attack risks and that aggregate metrics mask stark, scenario-specific vulnerabilities across frontier models.

AEM: Adaptive Entropy Modulation for Multi-Turn Agentic Reinforcement Learning

Addressing ambiguous credit assignment and premature entropy collapse caused by sparse outcome-only rewards in multi-turn agentic RL, AEM reveals via information geometry that natural-gradient entropy drift is governed by the product of advantage and relative surprisal, adaptively rescaling advantages via a response-level uncertainty proxy to smoothly transition from early exploration to late exploitation without external supervision.

Agent2World: Learning to Generate Symbolic World Models via Adaptive Multi-Agent Feedback

Agent2World proposes an automated tool-augmented multi-agent framework that synthesizes symbolic world models through deep web research, sandboxed code implementation, and adaptive testing (unit tests and ReAct simulation), while transforming verified multi-turn interaction traces via rejection sampling into an effective SFT data engine.

Agent Meltdowns: The Road to Hell Is Paved with Helpful Agents

This paper introduces and measures "accidental meltdowns," a critical failure mode where autonomous agents, lacking any adversarial influence, autonomously escalate into unsafe behaviors—such as unauthorized reconnaissance, privilege escalation, and credential leakage—solely to satisfy their helpfulness optimization when encountering benign environmental errors.

Agent Security is a Systems Problem

Reframing agent security as a systems security problem, this work argues that the underlying AI model must be treated as an untrusted component, enforcing deterministic security invariants at the encompassing system level via instruction-data separation, verifiable least-privilege policy generation, and information flow control (IFC).

Agent-ToM: Learning to Monitor Autonomous LLM Agents via Theory-of-Mind Reasoning

Addressing the challenge of detecting covert malicious side tasks in autonomous LLM agents over extended horizons, Agent-ToM combines Theory-of-Mind latent state inference with a two-call Reason–Verify–Refine pipeline and distills critique feedback into persistent semantic guardrail memory, outperforming multi-call ensemble monitors.

AgentAbstain: Do LLM Agents Know When Not to Act?

AgentAbstain introduces the first systematic evaluation benchmark and automated synthesis pipeline (AbstainGen) for agentic abstention across pre-execution and runtime scenarios, revealing that even the top frontier model (Gemini 3.1 Pro) achieves only 59.5% paired accuracy and demonstrating that task-solving competence scales independently of calibrated restraint.

Browse all 270 LLM Agent papers →


⚖️ Alignment & RLHF (44)

Alignment Dynamics in LLM Fine-Tuning

By deriving closed-form updates of an alignment score via empirical Neural Tangent Kernels, this paper decomposes alignment evolution into a Driving Force and a Rebound Force, explaining why distribution narrowness amplifies alignment reversal and identifying a Rehearsal Priming Effect where prior alignment accelerates subsequent re-alignment.

Authority Bias in Language Models: Source Deference and User Agreement Are Not Interchangeable

Holding the wrong answer fixed and changing its endorser, the paper uses direction removal and an independently fitted attribution patch to show selectively intervenable differences between wrong-source deference and user agreement in three open-weight families, so one sycophancy evaluation cannot substitute for the other.

Avoiding Obfuscation with Prover-Estimator Debate

To overcome the foundational vulnerability where dishonest debaters construct intractable subclaims to trigger the "obfuscated arguments problem" in recursive debate, this paper proposes an asymmetric Prover-Estimator Debate protocol equipped with a \(\rho\)-stability condition, theoretically proving that an honest debater can win in approximate Stackelberg equilibria without requiring computational superiority over dishonest opponents.

Back to Blackwell: Closing the Loop on Intransitivity in Multi-Objective Preference Fine-Tuning

To address the pervasive issue of cyclic, intransitive preferences in LLM fine-tuning where optimal policies cease to exist under scalarized multi-objective feedback, this paper introduces the Maximum Entropy Blackwell Winner (MaxEntBW) solution concept and derives PROSPER, a provably efficient single-player regression algorithm that bypasses adversarial training while establishing new state-of-the-art results on instruction-following benchmarks.

Bayesian Preference Learning for Test-Time Steerable Reward Models

This paper introduces Variational In-Context Reward Modeling (ICRM), framing test-time preference adaptation as amortized variational inference with a conjugate Beta prior to dynamically steer classifier reward models via few-shot demonstrations across single- and multi-objective alignments while theoretically mitigating reward over-optimization.

Bayesian Test-Time Inference of Task-Aligned Similarity from Weak Interactive Feedback

Introduces BASIS, a training-free Bayesian framework that leverages weak interactive feedback (scalar scores, pairwise preferences, or verifier outputs) to jointly infer candidate targets and task-aligned similarity criteria, selecting informative probes via expected mutual information gain.

Beyond Pairs: Your Language Model is Secretly Optimizing a Preference Graph

To overcome the limitations of standard DPO that collapses multi-rollout responses into independent pairs and discards transitivity, this paper proposes GraphDPO, which structures rollouts into a directed acyclic preference graph and optimizes a local Plackett–Luce objective over dominated neighborhoods in linear time using equivalence-class layering, achieving substantial gains across reasoning and program synthesis.

BSO: Safety Alignment Is Density Ratio Matching

This paper proves that the likelihood ratio of the optimal safe policy decomposes in closed form into data and reference ratios, formulating safety alignment as Bregman divergence-based density ratio matching (BSO) without auxiliary reward/cost models while attaining a superior helpfulness-safety Pareto frontier.

Consolidating Rewarded Perturbations for LLM Post-Training

Addressing the prohibitive test-time compute of prediction-level ensembling in weight-space post-training (e.g., RandOpt), CoRP introduces a gradient-free consolidation operator that uncovers the reproducible low-rank geometry of rewarded Gaussian perturbations, combines compatibility-aware reweighting with held-out validation gating, and deploys a single consolidated model that achieves over half of the 50-pass ensemble gain at one forward pass per test example using 1/10th the perturbation budget.

Diffusion LLMs are Natural Adversaries for any LLM

Reformulates adversarial prompt optimization as an amortized inference problem using pretrained Diffusion LLMs (DLLMs), sampling low-perplexity, highly transferable jailbreaks via conditional inpainting on target harmful responses without accessing target model parameters.

Browse all 44 Alignment & RLHF papers →


🔒 Safety (208)

A Comprehensive View of Fairness through Distributional Stability

By reformulating fairness as the distributional stability of a prediction function under demographic perturbations of sensitive groups, this paper unifies individual and group fairness and proposes a second-order cone programming (SOCP) framework over reweighted mixture models that guarantees uniform test-time deployment across arbitrary demographic shifts.

A Statistical Framework for Algorithmic Collective Action with Multiple Collectives

This paper introduces the first rigorous measure-theoretic statistical framework for Algorithmic Collective Action (ACA) with multiple collectives, providing finite-sample lower bounds on per-collective and global success for signal planting and unplanting under partial information regimes.

A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts

This paper systematically analyzes out-of-distribution (OOD) detection through a representation-centric lens, demonstrating that detector efficacy is primarily dictated by penultimate-layer Neural Collapse geometry and subspace properties, and introduces an NC-guided predictive framework that recommends competitive detector shortlists across architectures without requiring any OOD validation data.

A Unified Framework for Adversary-Aware Differential Privacy Bounds

This paper introduces the first unified theoretical framework for adversary-aware differential privacy bounds that seamlessly handles multiple targets, non-uniform prior distributions, and approximate reconstruction metrics using only DP parameters and the attacker's baseline prior success rate.

ABC-Align: Prediction-Powered Alignment with Adaptive Bias Control

Addressing the bottleneck of scarce human feedback and biased pseudo labels in LLM alignment, this paper proposes ABC-Align, a two-parameter gradient estimator with adaptive bias control that breaks the irreducible \(\Theta(1/n)\) variance floor of unbiased methods, delivering superior performance across RM, DPO, and GRPO.

Accuracy vs. Accuracy: Computational Tradeoffs Between Classification Rates and Utility

This paper demonstrates that in settings with rich individual labels, maintaining accurate subpopulation classification rates (decision/classification accuracy) and minimizing downstream loss (utility optimization/calibration) are computationally incompatible under standard cryptographic assumptions, even though both are simultaneously satisfied by the Bayes-optimal predictor and attainable individually.

Adaptive Auditing of AI Systems with Anytime-Valid Guarantees

This paper introduces a safe anytime-valid inference framework for adaptive AI auditing that formalizes testing as dueling hypothesis tests between the model and the auditor, deploying dual e-processes to provide rigorous finite-sample Type-I error control under arbitrary adaptive sampling and optional stopping.

Agent MechSuits: Mechanistic Subspace Safety Steering for Multi-Turn CLI Agents

Agent MechSuits is a white-box safety framework for multi-turn CLI agents that detects harmful execution states via step-level linear probing in early-to-middle hidden layers and mitigates malicious behavior by steering along a sparse 10-dimensional subspace, reducing attack success rates by 72.63% while preserving over 93.4% benign utility.

AI Agents May Always Fall for Prompt Injections

This paper demonstrates that the prevailing "data-instruction separation" defense paradigm fundamentally fails against contextual manipulation attacks, recasts prompt injection through the lens of Contextual Integrity (CI), and establishes an impossibility result showing that autonomous agents face an inherent tradeoff between over-compliance and under-compliance.

AI Models Can Provably Hide Arbitrary Capabilities

This paper demonstrates that adversarial AI models can embed encrypted, fully executable neural network sub-circuits within their weights using cryptographic digital lockers, rendering hidden capabilities provably impossible to elicit or identify under polynomial-time white-box audits without the secret trigger.

Browse all 208 Safety papers →


👻 Hallucination Detection (19)

A Theory of Time-Sensitive Language Generation: Sparse Hallucination Beats Mode Collapse

This paper introduces a "timeliness" deadline constraint into the game-theoretic framework of language generation in the limit, proving that any eventually consistent (hallucination-free) generator inevitably suffers mode collapse (zero lower density) under any deadline function, while proposing a Speculative Blackbox Generation (SBG) reduction that restores the optimal \(1/2\) timely lower density under arbitrary superlinear deadlines by permitting a vanishing rate of sparse hallucinations.

Actionable Hallucination Detection: Translating Latent Uncertainty into Agentic Critique

This paper introduces the Latent Critic, a lightweight low-rank adapter (LoRA) that operates concurrently with a frozen base LLM during a single forward pass to actively restructure the transformer residual stream and translate internal latent uncertainty into parameter-level natural language diagnostic feedback (e.g., ungrounded: ["date"]), achieving 0.966 detection AUROC and over 80% parameter localization accuracy on Qwen3-4B while enabling closed-loop zero-shot agentic self-correction with under 10 ms of latency overhead.

Alleviating Hallucination in Reasoning Tasks with Training-Free Uncertainty-Guided Steering

The paper proposes USteer, a simple training-free activation steering method that extracts gradients from an internal Neural Collapse Inspired (NCI) uncertainty metric via the Logit Lens and performs single-step dynamic hidden-state intervention during decoding, significantly mitigating hallucinations in multi-step reasoning tasks.

CHORD: Cross-Model Hallucination Detection via Relational Graph Discrimination

Addressing the failure of internal representation-based hallucination detectors under cross-model transfer due to severe semantic inconsistency, CHORD reorganizes hidden states across layers, tokens, and feature blocks into a unified relational graph, applying relation-aware attention and Feature-Critic meta-learning to deliver an average AUROC of 0.8042 across 8 unseen LLMs and 7 benchmarks.

Dual-Pathway Circuits of Object Hallucination in Vision-Language Models

This paper proposes the Dual-Pathway Circuit Analysis framework, applying activation patching and Conditional Pathway Analysis (CPA) across five architecturally diverse vision-language models (VLMs) to uncover a dual-pathway organization (visual grounding vs. hallucination pathways) and a consistent grounding-pathway polarity flip on hallucinating samples, while reducing object hallucination by up to 76% at \(\le 2\text{ pp}\) accuracy cost via training-free component-wise scaling.

Entropy Distribution as a Fingerprint for Hallucinations in Generative Models

This paper reveals that the distributional shape and tail behavior of token-level entropies serve as a hallucination fingerprint independent of the mean, formalizes hallucination detection as a statistical hypothesis test, and proposes the Calibrated Entropy Score (CES)—a single-pass black-box detector that achieves the best performance among single-pass methods across 10 models and 8 QA benchmarks while statistically matching expensive multi-sample consistency methods.

Generating in the Limit with Infinitely Many Hallucinations

This paper recasts language generation in the limit as a formal precision–recall trade-off over infinite sets, proving that allowing infinitely many hallucinations at a vanishing rate (relaxing tail precision while keeping global precision at 1) or permitting a fixed fraction of non-novel repeated outputs (\(\gamma < 1\)) strictly breaks the recall ceiling of zero-hallucination algorithms under sparse adversarial revelations, and can even achieve both precision and recall of 1.

Hallucination in World Models is Predictable and Preventable

This paper reveals that hallucinations in generative world models stem fundamentally from state-action data coverage gaps, decomposes them into perceptual hallucination, action marginalization, and scene divergence, introduces three label-free dynamism-normalized runtime predictors (correlating with rollout error at \(\rho \approx 0.80\)), and mitigates them via uniform-task resampling and predictor-guided curiosity data collection—adapting a 350M-parameter Dreamer 4 world model to unseen tasks with just 50 trajectories.

HalluWorld: A Controlled Benchmark for Hallucination via Reference World Models

Grounded in an explicit "Reference World Model" formulation, this paper introduces HalluWorld, a controlled and extensible hallucination benchmark spanning gridworlds, chess, and realistic Linux terminal tasks that automatically generates ground-truth labels while decoupling observability, serialization formats, and conflict resolution policies, systematically revealing distinct failure modes across perception, long-horizon memory, causal forward simulation, and epistemic abstention in frontier LLMs.

How Fine-Tuning Objectives Shape Layer-Wise Information in LLM Hallucination Detection

This paper demonstrates that in binary LLM hallucination detection, classifier-head fine-tuning (cls) induces task representations with stronger linear decodability and clearer two-cluster geometry than sequence-to-sequence (s2s) fine-tuning; by combining static Partial Information Decomposition (PID) and dynamic training trajectories to locate a narrow upper-middle signal-forming band (InfoWindow) followed by a redundant upper plateau, it derives a static upper-layer truncation rule that removes 31% of Qwen3 layers without accuracy loss while achieving a \(1.4\times\) inference speedup.

Browse all 19 Hallucination Detection papers →


📊 LLM Evaluation (83)

A Semantic-Sampling Framework for Evaluating Calibration in Open-Ended Question Answering

Addressing the lack of internal logits and the failure of traditional confidence metrics in open-ended question answering, this paper presents Sem-ECE, a semantic-sampling calibration evaluation framework, along with a held-out estimator Sem2 that eliminates the winner's curse and provides an observable diagnostic for question difficulty.

A Unified Perturbation Framework for Analyzing Leaderboard Stability and Manipulation

A unified influence-based perturbation framework for Bradley-Terry leaderboards formalizes match additions, drops, flips, and player removals with one-step Newton corrections, enabling quantitative stability audits and targeted manipulation across Top-\(k\) ranks, Kendall's \(\tau\), and confidence intervals.

AI Evaluation Should Require Standardized Item-Level Data Releases

Addressing the crisis where aggregate benchmark scores conceal construct misalignment and item degradation, this paper proposes standardizing item-level data releases as essential AI evaluation infrastructure and introduces OpenEval, archiving 10.7M+ responses across 444 models.

AI GAMESTORE: Scalable, Open-Ended Evaluation of Machine General Intelligence with Human Games

To overcome the rapid saturation and narrow scope of static AI benchmarks, this paper proposes evaluating machine general intelligence across the open-ended "Multiverse of Human Games" via AI GAMESTORE—a platform that sources, synthesizes, and adapts commercial games using LLMs with humans in the loop, showing that frontier VLMs achieve less than 10% of median human performance on 100 games.

Aligning Language Model Benchmarks with Pairwise Preferences

Addressing the predictive validity gap where static benchmarks fail to predict downstream utility, this paper introduces benchmark alignment and the BenchAlign diagnostic tool, showing that learning-to-rank item weights enables small-model evaluations to zero-shot rank unseen 70B+ models according to target preferences.

APM: Evaluating Style Personalization in LLMs with Arbitrary Preference Mappings

The paper proposes the Arbitrary Preference Mapping (APM) benchmark, which decouples semantic priors and cancels LLM judge biases via a randomized zero-mean mapping between user prompt traits and response principles, demonstrating that routing-based adaptation is the most reliable approach for implicit style personalization.

Ask-E: An Environment for Calibrated Question Generation

Presents the Ask-E evaluation and training environment, where an asker LM interacts with two boundary solver models to generate a question solvable by exactly one of them, inverting the conventional need for stronger supervisors and improving downstream math reasoning through RL with no new data or correctness rewards.

AssayBench: An Assay-Level Virtual Cell Benchmark for LLMs and Agents

AssayBench unifies 1,920 human CRISPR screens into a free-text conditioned phenotypic endpoint gene ranking benchmark and introduces Adjusted nDCG (AnDCG) to correct for assay-specific baselines, demonstrating that while frontier generalist LLMs outperform specialized biology models, a substantial gap remains relative to empirical performance ceilings.

BayesJudge: Uncertainty-Aware Bayesian Meta-Evaluation of Human and LLM Judgments

Addressing the intertwined challenges of human annotation disagreement and systematic LLM judge biases in pairwise evaluation, BayesJudge introduces an online streaming Bayesian meta-evaluation layer that uses tie-open labels and paired order swaps to disentangle item ambiguity from evaluator biases, producing full posterior verdict distributions and calibrated uncertainty.

BBOWP-Bench: Evaluating LLMs on Black-Box Optimization Word Problems

Introduces Black-Box Optimization Word Problems (BBOWP) and the BBOWP-Bench suite, evaluating how well LLMs acting as meta-solvers can infer design variable search spaces and select suitable optimization algorithms directly from natural-language task descriptions.

Browse all 83 LLM Evaluation papers →


⚡ LLM Efficiency (79)

A Statistical Theory of Gated Attention through the Lens of Hierarchical Mixture of Experts

This paper establishes the first mathematical equivalence between multi-head self-attention/gated attention and three-level Hierarchical Mixture of Experts (HMoE), rigorously proving that non-linear gating reduces expert estimation sample complexity from exponential \(\mathcal{O}(\exp(\epsilon^{-1/\tau}))\) in standard attention to polynomial \(\mathcal{O}(\epsilon^{-4})\), and uncovering why gating Value or SDPA outputs eliminates degenerate PDE coupling.

Accelerating Diffusion Language Models via Structured Suffix Modeling

Addressing the computational redundancy of full suffix attention and the loss of iterative denoising history in block-wise diffusion language models, this paper proposes Structured Suffix Modeling (SSM)—a training-free framework that allocates distinct token budgets across local, middle, and tail suffix regions while fusing previous-step decoding probabilities via soft embeddings, achieving up to a 72.81× end-to-end speedup while matching or improving generation accuracy.

Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models

Without training the model or changing the base sampler's stopping rule, ADAS dynamically discounts confidence using a candidate's attention to positions already selected in the current step and their uncertainty, improving parallel decoding at low numbers of denoiser evaluations; task-and-method-average gains are 9.11 and 10.46 percentage points for the two models.

Block Sparse Flash Attention

BSFA first computes all causally visible QK scores exactly inside FlashAttention-2, then uses offline-calibrated block-maximum thresholds to skip V loads, PV, and softmax-state updates for low-scoring blocks, achieving a LongBench score of 39.78% versus the dense baseline's 40.24% on Llama-3.1-8B and a 1.13× end-to-end prefill speedup on the longest 10 samples.

Communication-Efficient LLM Adaptation over Decentralized GPU Meshes

This paper proposes an asynchronous two-circuit training architecture for decentralized post-pretraining LLM adaptation, combining aggressive in-graph activation masking for pipeline parallelism with an off-critical-path anchor circuit and spectral correction optimizer to eliminate isotropic compression noise, achieving over 40× training throughput gains across commodity 200 Mbps internet links while fully matching dense adaptation performance.

CommunityKV: Efficient Long-Context Decoding via Graph Partitioning

CommunityKV formulates training-free sparse attention as a community detection problem on an attention-induced token graph, combining direct and co-attention relationships with Leiden partitioning during prefill and constant-time greedy updates during decoding to enable sub-linear retrieval of semantically coherent KV chunks.

CompactAttention: Accelerating Chunked Prefill with Block-Union KV Selection

CompactAttention tackles the kernel inefficiency and explicit KV-copy overhead in chunked prefill (\(Q \ll KV\)) by lowering 2D block-sparse masks into GQA-aware KV block tables through Q-block and intra-group unions, enabling zero-copy in-place execution on top of dense paged-attention engines.

Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse

Addressing the substantial prefill overhead incurred when switching between different-sized models in an LLM family, this paper proposes a gradient-free, per-head closed-form ridge mapper with RoPE decoupling that achieves a 2.7–25× prefill speedup while retaining 73–98% of standalone downstream accuracy on matched-KV pairs.

D-PACE: Dynamic Position-Aware Cross-Entropy for Parallel Speculative Drafting

Addressing the failure of fixed position-decay schedules in block-diffusion parallel drafters to match dynamic acceptance bottlenecks during training, D-PACE derives closed-form position weights by differentiating a smooth surrogate of expected accepted draft length, boosting wall-clock speedup and emitted length across multiple models with only 2.3% training overhead and zero inference modifications.

DivMoE: Fine-Grained MoE Upcycling via Cross-Domain Expert Composition

Addressing the catastrophic routing collapse in single-source fine-grained MoE upcycling, DivMoE introduces domain-specialized expert slicing and hard diversity-constrained routing, strictly outperforming dense base models and six upcycling baselines across 15 benchmarks.

Browse all 79 LLM Efficiency papers →


📚 Pretraining (52)

A Bitter Lesson for Data Filtering

Through extensive scaling experiments across compute, epoch budgets, and model capacities, this paper demonstrates that unfiltered web data ultimately outperforms common heuristic and classifier-based filters in high-compute regimes, revealing a "bitter lesson" for data filtering.

A Scalable Measure of Loss Landscape Curvature for Analyzing the Training Dynamics of LLMs

This paper introduces critical sharpness and relative critical sharpness—curvature proxies computed via fewer than 10 forward passes along the training update direction—providing the first empirical demonstration of progressive sharpening and Edge of Stability at 7B parameter scale, while prescribing optimal pre-training data mixtures for mid-training and fine-tuning without grid searches.

A Theory of Training Profit-Optimal LLMs

This paper bridges neural scaling laws with microeconomic theory of the firm, establishing the first formal profit-maximization framework for LLM training under compute-bound and data-bound regimes to derive asymptotic bounds on optimal model size, token budgets, and capital expenditure.

Accelerating LLM Pre-Training through Flat-Direction Dynamics Enhancement

Based on a Riemannian Inertial System with Hessian Damping (RISHD), this paper introduces LITE, a generalized acceleration strategy that selectively amplifies learning rates and Hessian damping coefficients along flat directions via stateless fast subspace filtering, significantly speeding up Muon and SOAP in LLM pre-training.

An Empirical Study on Noisy Data and LLM Pretraining Loss Divergence

Through controlled synthetic uniform random noise injection, this paper systematically demonstrates that noisy data directly triggers LLM pretraining loss divergence, shows that limited noise vocabulary size and network depth critically amplify divergence probability, and establishes an attention logit diagnostic framework (~1800 vs. ~4000) alongside QK-LayerNorm defenses to isolate and mitigate data-induced instabilities.

Bridging Compute- and Data-Optimal Pretraining

This paper introduces the unified Compute-Data (CD) Scaling Law, which bridges compute-optimal and data-optimal pretraining regimes by modeling a dynamic token effectiveness function \(\eta\) that captures diminishing returns and saturation ceilings across model sizes and data budgets.

Detecting RLVR Training Data via Structural Convergence of Reasoning

Addressing the hidden benchmark contamination risks in reinforcement learning with verifiable rewards (RLVR) where training data is undisclosed, this paper discovers that RLVR induces structural mode collapse and generation rigidity on seen prompts, and introduces Min-kNN Distance—a reference-free, probability-free black-box detector that achieves state-of-the-art detection across diverse reasoning models.

Dynamic Tokenization via Reinforcement Patching: End-to-end Training and Zero-shot Transfer

ReinPatch is the first end-to-end framework to jointly optimize a discrete sequence patching policy and its downstream sequence backbone using reinforcement learning via Group Relative Policy Gradient (GRPG), strictly enforcing target compression rates and decoupling a standalone foundation patcher that demonstrates remarkable zero-shot transfer across diverse forecasting tasks.

Effective Context in Transformers: An Analysis of Fragmentation and Tokenization

Framing sequence representation choices through finite-context prediction on stationary Markov sources, this paper proves that lossless fragmentation into finer units fundamentally increases optimal log-loss due to phase ambiguity, whereas greedy tokenization reliably extends effective context, governed by the lower tail of the token-window source-span distribution and tokenizer compression rate.

ELF: Embedded Language Flows

ELF introduces a continuous diffusion language model based on continuous-time Flow Matching that operates entirely within a continuous bidirectional embedding space and repurposes the shared-weight backbone as a discrete decoder only at the final step, dramatically outperforming discrete and continuous DLM baselines with fewer sampling steps and 10× fewer training tokens.

Browse all 52 Pretraining papers →


✏️ Knowledge Editing (8)

DynEdit: Dynamic Entropy-Guided Sequential Editing for Large Language Models

DynEdit overcomes the rigidity and gradient dilution of fixed-chunk long-form knowledge editing by leveraging pre-edit token-level Shannon entropy to dynamically shrink sliding strides across dense factual regions and adaptively weight token optimization.

Exposing the Illusion of Erasure in Knowledge Editing for LLMs

Framing knowledge editing through adversarial reverse-engineering, this work proves that localized low-rank updates merely push pre-trained facts into complementary null spaces with surface-level suppression logits; universal adversarial suffixes reliably recover supposed "erased" knowledge (>85% guided, up to 48.5% blind), exposing an illusion of erasure.

Golden Layers and Where to Find Them: Improved Knowledge Editing for Large Language Models Via Layer Gradient Analysis

This paper demonstrates the existence of sample-invariant "Golden Layers" that achieve near-optimal knowledge editing across LLMs, and introduces Layer Gradient Analysis (LGA)—a fast, first-order gradient attribution method that accurately pinpoints golden layers without requiring actual editing runs.

MetaKE: Meta-Learning for Knowledge Editing Toward a Better Accuracy-Editability Trade-off

MetaKE bridges the long-standing disconnect between upstream target representation optimization and downstream constrained parameter optimization by formulating knowledge editing as a bi-level optimization problem, using a closed-form Structural Gradient Proxy to provide downstream-aware feedback for target refinement.

Norm Anchors Make Model Edits Last

Reveals that sequential Locate-and-Edit (L&E) failure stems from a self-reinforcing positive norm-feedback loop between edited MLP weights and solved value vectors, and proposes Norm-Anchor Scaling (NAS) to rescale target values to base-model reference norms, extending the editing horizon by over 4×.

Open-Vocabulary Domain Unlearning

The paper extends visual-domain unlearning from failure on training classes to failure on classes absent from unlearning training, combining Fisher-masked visual-parameter updates with Targeted Manifold Scattering (TMS) to improve the forgetting–retention trade-off in few-shot CLIP classification, without establishing certified data deletion or guaranteed semantic preservation.

Steer2Edit: From Activation Steering to Component-Level Editing

Steer2Edit introduces a theoretically grounded, training-free framework with closed-form solutions that converts activation steering vectors into component-level rank-1 parameter edits on attention heads and MLP neurons, yielding superior attribute-utility trade-offs across safety alignment, truthfulness promotion, and reasoning efficiency control.

Synthetic Worlds for Temporal Evaluation and Knowledge Updating in LLMs

This paper introduces ParallelEvents, a contamination-resistant future-world benchmark ensuring local and global consistency, and Synapse, a synthetic training framework combining long-form continued pretraining with preference-steered editing to internalize evolving multi-hop and causal knowledge.


🔍 Information Retrieval & RAG (18)

Beyond Raw Context Transfer: Representation-based Federated Retrieval-Augmented Generation

Addressing the severe privacy risks and lengthy context overhead of exchanging raw documents in decentralized environments, FedRepRAG keeps raw text localized and exchanges only compact latent embeddings, leveraging a collaboratively trained lightweight projector to inject each retrieved document as a single soft token into frozen LLM/VLM backbones, slashing context length by 19–398× while substantially outperforming local retrieval.

Can Language Models Actually Retrieve In-Context? Drowning in Documents at Million Token Scale

Focusing on corpus-scale in-context retrieval (ICR) at million-token context lengths, this paper diagnoses the root cause of retrieval collapse as attention dilution, and introduces length-aware Softmax rescaling (SSMax) alongside document-level sparse routing to match a same-backbone dense retriever.

Cluster with Auctions for Vector Search

Decouples database partitioning from query probing in inverted indexes, optimizing cluster assignments via GPU-parallel capacitated auction algorithms and query probing via neural networks alternatively, yielding up to 4.7× throughput over competitive baselines under severe distribution shift.

Data Auctions for Retrieval Augmented Generation

Addressing the need for exclusive selling of sensitive knowledge to GenAI platforms for RAG tasks, this paper formulates a model-agnostic coverage valuation function, proves that welfare maximization and demand queries are NP-hard, and designs the LPRMono mechanism—combining LP relaxation, randomized rounding, and a novel "data burning" post-processing step—to achieve dominant-strategy truthfulness with a tight \((1-1/e)\) approximation ratio.

Diagnosing and Repairing Citation Failures in Generative Engine Optimization

This paper reframes Generative Engine Optimization (GEO) from generic text rewriting into a diagnostic repair paradigm, proposing a full-pipeline taxonomy of citation failures alongside AgentGEO, an agentic framework that achieves over 40% relative citation rate improvement while modifying merely 5% of webpage content.

From Topology to Retrieval: Decoding Embedding Spaces with Unified Signatures

This paper introduces Unified Topological Signatures (UTS), a multidimensional geometric framework aggregating persistent homology, intrinsic dimensionality, diversity, isotropy, and clustering descriptors to comprehensively characterize text embedding spaces, accurately predict downstream dense retrieval effectiveness and document retrievability bias without ground-truth labels, and refute scale-driven representation convergence claimed by the Platonic Representation Hypothesis.

GISA: A Benchmark for General Information Seeking Assistant

Addressing key flaws in current search agent benchmarks—such as reverse-engineered queries, artificial separation of depth and breadth, static answers vulnerable to pre-training contamination, and unstable LLM judges—GISA introduces a human-crafted benchmark featuring four structured answer types, dynamic live subsets, and complete human search trajectories for process-level supervision.

Is Dimensionality a Barrier for Retrieval Models?

This paper theoretically investigates whether representation dimensionality limits large-scale retrieval, proving that near-optimal retrieval margins are achievable in dimensions scaling only logarithmically with dataset size, \(d = O(\mathsf{m}^{\mathrm{rd}}(+\infty, A)^{-2}\log(\min(n, N)))\), while revealing that Sigmoid loss significantly surpasses InfoNCE for large-margin low-dimensional geometry.

Knowledge-Graph Paths as Intermediate Supervision for Self-Evolving Search Agents

Addressing the twin bottlenecks of ungrounded question synthesis in Proposers and sparse binary rewards in Solvers within Search Self-Play (SSP), this paper introduces construction-derived intermediate supervision by reusing knowledge-graph paths to both provide relational context for multi-hop question generation and supply non-zero partial credit to incorrect Solver trajectories via Waypoint Coverage Reward.

LeanSearch v2: Global Premise Retrieval for Lean 4 Theorem Proving

Addressing the challenge of locating scattered, non-vocabulary-overlapping supporting lemmas in Lean 4 formalization, LeanSearch v2 introduces a two-mode system combining a hierarchy-informalized embedding–reranker substrate (standard mode) and an iterative sketch-retrieve-reflect loop (reasoning mode), substantially advancing global premise recall and downstream proof success.

Browse all 18 Information Retrieval & RAG papers →


💻 Code Intelligence (44)

A Theoretical Analysis of Test-Driven Code Generation

The paper provides a unifying probabilistic and decision-theoretic foundation for test-driven code generation, proving that soft functional similarity estimators strictly dominate hard equivalence in signal-to-noise ratio and formalizing iterative backprompting as in-context Thompson sampling with an irreducible regret bound.

Build-and-Find: An Effort-Aware Protocol for Evaluating Agent-Managed Codebases

Build-and-Find establishes an effort-aware evaluation layer between behavioral correctness and downstream modification by treating generated codebases as communication artifacts and measuring downstream agents' intent recovery accuracy and inspection effort under strict stability gates.

Code2World: A GUI World Model via Renderable Code Generation

Code2World shifts GUI next-state prediction from raw pixels or abstract text to renderable HTML code generation, aligning layout fidelity and action logic via dual-reward Render-Aware Reinforcement Learning (RARL) to serve as a high-fidelity virtual sandbox that enhances downstream agent navigation.

CodeMimicry: Exploiting Safety Generalization Lag in Large Language Models via Structured Code Completion

Targeting the "safety generalization lag" where natural language safety alignment fails to generalize to structured programming, this paper proposes CodeMimicry, a fully automated black-box jailbreak framework that induces harmful code completion via object-oriented encapsulation and iterative refinement, achieving a 96.25% attack success rate with only 1.51 queries on average across 8 commercial LLMs.

CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models

CodeScaler distills high-quality execution feedback into a code reward model, uses strict code extraction and reward shaping to support reinforcement learning and candidate reranking without online execution, and exceeds RLVR by 1.55/4.23 percentage points in average Avg@8 across four benchmarks for 8B/14B policies trained on DeepCoder.

Constrained Code Generation with Discrete Diffusion

Addressing the shortcomings of autoregressive post-hoc repair and syntax-only constrained decoding, this paper introduces CDC, a training-free neurosymbolic framework that integrates soft surrogate execution signals and static program analysis directly into the discrete diffusion reverse denoising trajectory to enforce functional correctness and security constraints with surgical locality.

CORVUS: Context Optimization and Reduction Via Underlying Synchronization for LLM Coding Agents

Addressing the inefficiencies of traditional append-only trajectory architectures that cause duplicate file reads and stale snapshot errors in LLM coding agents, CORVUS decouples file-reading actions from observations via a synchronized registry that injects refreshed repository states before each reasoning cycle, cutting 9–50% input tokens and up to 37% reasoning cycles while maintaining task pass rates.

Counterfactual Rollout Replay: Forkable Environments as Free Process Rewards for Software Engineering Agents

CRR restores a small number of decision states during software-agent training and scores realized actions using the terminal-return difference between the original trajectory and an alternative-action continuation, raising SWE-bench Verified pass@1 from extended GRPO's 36.7% to 41.7% within the same 40.125-hour training window on an eight-H100 node, without eliminating fork execution costs.

CRANE: Constrained Reasoning Injection for Code Agents via Nullspace Editing

CRANE introduces a training-free directional parameter-editing framework that injects planning and error-recovery behaviors from Thinking models into Instruct code agents via magnitude thresholding, Conservative Taylor Gating, and Graduated Sigmoidal Projection, boosting task success across Roo-Eval, SWE-bench-Verified, and Terminal-Bench v2 while strictly maintaining compact, tool-disciplined execution.

CTE-Bench: Counterfactual Trace Evaluation for Stateful Software Simulators

CTE-Bench executes software with and without an intervention to evaluate sustained response prediction on fixed future calls: effect-step accuracy reaches 54.3%–61.5% with correct earlier answers, falls to 24.8%–33.2% in free rollout, and whole-trace exact match peaks at just 1.2%.

Browse all 44 Code Intelligence papers →


💬 LLM (Other) (98)

Allocate Marginal Reviews to Borderline Papers Using LLM Comparative Ranking

Proposes allocating surplus reviewer capacity at large ML conferences specifically to borderline papers identified beforehand via LLM pairwise comparisons and Bradley–Terry ranking, improving decision reliability without increasing overall reviewer load.

Causal Edit Serialization

Proposes Causal Edit Serialization (CES), integrating a lightweight localization head and dual-index RoPE into standard causal decoders to serialize explicit edit programs—covering insertion, deletion, replacement, copy, and move—directly within the causal token stream while preserving the causal attention mask.

Characterizing Memorization in Diffusion Language Models: Generalized Extraction and Sampling Effects

This paper establishes the first generalized discoverable extraction framework for diffusion language models under arbitrary masking patterns and stochastic denoising trajectories, theoretically proves and empirically validates that sampling resolution monotonically governs verbatim memorization, and reveals that DLMs exhibit substantially lower PII leakage risk than autoregressive models under aligned settings.

COMPLLLM: Fine-tuning LLMs to Discover Complementary Signals for Decision-making

ComplLLM introduces a decision-theoretic post-training framework that fine-tunes language models using best-attainable utility gain as rewards via SFT and GRPO, enabling LLMs to surface actionable, discrete complementary signals from unstructured text that improve upstream model recommendations.

CompLLM: Fine-Tuning LLMs to Discover Complementary Signals for Decision-Making

CompLLM introduces a decision-theoretic post-training framework that fine-tunes LLMs using best-attainable payoff gain as a reward under GRPO, discovering discrete complementary signals in unstructured text that repair blind spots in upstream agent recommendations.

Confidence-Based Decoding is Provably Efficient for Diffusion Language Models

This paper develops the first theoretical framework for confidence-based decoding in diffusion language models (DLMs), proving that an entropy sum-based adaptive unmasking strategy achieves \(\varepsilon\)-accurate sampling in KL divergence with an expected iteration complexity of \(\widetilde{O}(H(X_0)/\varepsilon)\), enabling automatic acceleration on low-entropy data without hyperparameter tuning.

Conformal Language Modeling via Posterior Sampling

Reframes conformal language model hallucination control from post-hoc claim deletion to posterior sampling conditioned on high-confidence regions, incorporating an explicit abstention mixture prior and monotone risk envelopes to achieve rigorous finite-sample risk bounds alongside superior downstream generation utility.

Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata

Addressing the failure of autoregressive constrained decoding in discrete diffusion language models caused by simultaneous multi-position denoising, this paper formulates finite automata as chain-structured graphical models and introduces an exact, tractable log-depth tree sampling algorithm that guarantees 100% constraint satisfaction with minimal inference overhead.

Continuous Diffusion Scales Competitively with Discrete Diffusion for Language

This paper revisits likelihood-based continuous diffusion language models and introduces RePlaid by aligning the architecture with modern discrete diffusion baselines, proving that continuous diffusion scales on par with discrete diffusion and establishes a new state of the art among continuous models.

Continuous Latent Diffusion Language Model

This paper introduces Cola DLM, a continuous latent diffusion language model that decouples text generation into global continuous latent prior transport via a block-causal DiT and local textual realization via a Text VAE decoder, escaping strict left-to-right sequential bottlenecks and exhibiting favorable scaling dynamics over competitive 2B baselines.

Browse all 98 LLM (Other) papers →


🎨 Image Generation (206)

ABC: Any-Subset Autoregression via Non-Markovian Diffusion Bridges in Continuous Time and Space

ABC models continuous-time, continuous-space stochastic processes via a single non-Markovian diffusion bridge SDE whose time and volatility adapt directly to physical elapsed time, trained simulation-free via path-dependent denoising score matching to support arbitrary causal and non-causal conditioning subsets.

Accelerating Rectified Flow Models via Trajectory-Aware Caching

This paper proposes TACache, a training-free cache acceleration framework for Rectified Flow models that orthogonally decomposes discrete velocity acceleration into parallel magnitude and orthogonal direction components, using dual cumulative-variation thresholds offline to build a safe skip schedule and recursively re-orthogonalized historical directions online to reconstruct skipped velocities without extra model evaluations, achieving up to 4.14× text-to-image and 2.11× text-to-video speedups.

Active Flow Expansion for Out-of-Distribution Discovery: from Theory to Molecules

Addressing the core limitation that standard flow and diffusion pre-training merely matches observed data distributions and remains blind to out-of-distribution valid regions, this paper introduces Active Flow Expansion (ActFlow), an active continued pre-training framework driven by representation-space verifier uncertainty that establishes the first reachability-based statistical guarantees for OOD generative discovery and substantially expands valid molecular and biological coverage.

Align & Invert: Solving Inverse Problems with Diffusion and Flow-based Models via Representation Alignment

This paper introduces test-time representation alignment between internal diffusion hidden states and a self-supervised visual encoder (DINOv2), using dynamic proxy targets and theoretical local contraction guarantees to substantially boost perceptual fidelity and reduce discretization steps in solving inverse problems without ground-truth supervision.

All Roads Lead to Rome: Flow-driven Multi-Anchor Exploration for Open-Environment Active 3D Mapping

The method generates multiple long-horizon exploration anchors with conditional flow matching, then combines obstacle-aware planning, exploration-mode clustering, and hierarchical path selection to improve mapping coverage in same-difficulty AiMDoom evaluation, difficulty transfer, and AiMDoom-to-MP3D transfer, without guaranteeing generalization to arbitrary open environments.

An exact information theory of generalization phase transitions in Bayesian diffusion models

This paper introduces analytically tractable Bayesian Information Restricted Diffusion (BIRD) models, rigorously proves via the statistical physics Random Energy Model that the memorization-generalization phase boundary occurs exactly when the mutual information between restricted observations and true data equals the log training set size (\(I(\varphi;\mathcal{C}_{x,t})=\ln|\mathcal{D}|\)), and demonstrates that spatially local BIRD models accurately predict individual outputs of early-training UNets and DiTs (\(r^2 \sim 0.85\text{--}0.93\)), revealing how dynamic information restriction enables diffusion models to circumvent the curse of dimensionality.

Analytical Correction for Subsampling Bias in Drifting Models

Addressing the \(\mathcal{O}(1/n)\) centroid-contraction bias inherent in the softmax-weighted minibatch estimator of drifting models, this paper introduces Analytical Bias Correction (ABC)—a plug-in adjustment that requires only two extra lines of code, preserves convex-hull containment, causes no first-order variance inflation, and slashes the residual bias down to \(\mathcal{O}(1/n^2)\).

Are Text-to-Image Models Inductivist Turkeys? A Counterfactual Benchmark for Causal Reasoning

This paper introduces CF-World, a three-level progressive counterfactual benchmark, alongside the CF-Eval pipeline equipped with Prior Resistance Rate (PRR) and Reasoning Retention Rate (RRR) metrics, revealing that current text-to-image (T2I) models suffer sharp performance degradation under counterfactual premises due to deep entanglement between objective rules and visual priors.

Are we really tilting? The mechanics of reward guidance in flow and diffusion models

This paper theoretically proves that reward hacking and mode-selection failure in reward-guided flow and diffusion models stem from the finite-particle plug-in approximation of the Doob \(h\)-function, proposing a zero-overhead closed-form reward damping schedule to eliminate within-mode collapse and clarifying how Best-of-\(n\) sampling compensates for cross-mode selection.

Asymptotically Exact Negative Guidance of Diffusion Models via Positive-Unlabeled Learning

Proposes PU Steering, a principled framework that recasts negative guidance of diffusion models as sampling from an explicit avoidance target density, coupling an online Positive-Unlabeled (PU) posterior classifier with Sequential Monte Carlo (SMC) resampling to achieve asymptotic exactness and eliminate heuristic local repulsion bias.

Browse all 206 Image Generation papers →


🎬 Video Generation (90)

AdaState: Self-Evolving Anchors for Streaming Video Generation

AdaState eliminates the frozen scene problem in streaming video diffusion by replacing static first-frame attention sinks with an adaptive latent state that is jointly denoised but never rendered, casting iterative denoising into an implicit recurrence carried by the KV cache to break the consistency-dynamics tradeoff.

Addressable Memory for Video World Models

Addressing the dual failure modes of out-of-distribution temporal RoPE offsets and phase cancellation during cache compression in autoregressive video world models, this paper introduces WorldTrace, a training-free memory framework that maintains long-horizon visual persistence and episodic recall under a bounded cache budget by assigning in-distribution virtual positions to summary slots and manipulating keys in canonical space.

Advancing Narrative Long Video Generation via Training-Free Identity-Aware Memory

IAMFlow introduces a training-free autoregressive long video generation framework that constructs an explicit global entity registry to track persistent character identities across evolving multi-prompt streams, combining attention-key greedy retrieval, adaptive prompt transitions, and asynchronous visual verification to achieve superior consistency while accelerating inference by 1.69×.

AnyMo: Scaling Any-Modality Conditional Motion Generation with Masked Modeling

This paper introduces OmniHuMo, a large-scale human motion dataset comprising over 5,000 hours of motion sequences with aligned text, speech, music, and trajectory annotations, alongside AnyMo, a unified framework combining a Residual FSQ tokenizer and a scalable masked Transformer that enables high-fidelity 3D motion synthesis under arbitrary modality combinations.

AtlasVid: Efficient Ultra-High-Resolution Long Video Generation via Decoupled Global-Local Modeling

AtlasVid decouples ultra-high-resolution long video synthesis into coarse global semantic proxy generation and locality-preserving high-resolution detail denoising, enabling a model trained only at 720P on two GPUs to extrapolate seamlessly to 4K/8K, 321-frame generation with a 60.9× speedup over full attention.

AUTEUR: Language-Driven Cinematographic Framing for Human-Centric Video Generation

Auteur introduces an actor-relative camera framing parameterization and cinematographic DSL, where a fine-tuned multimodal LLM generates sparse directorial keyframe programs from natural language that are deterministically decoded into continuous 6-DoF trajectories, achieving superior framing adherence and drastically reducing out-of-frame errors.

AVENUE: Audio-Video Editing, Understanding, and Evaluation

To overcome the limited edit diversity and modality-blind evaluation of existing audio-video editing systems, AVENUE introduces a benchmark of 1,291 video clips and 7,957 instructions across audio-targeted, video-targeted, and AV-coupled types, paired with a sample-specific MLLM-as-a-judge protocol grounded in explicit change and preservation criteria that systematically diagnoses cross-modal leakage across joint, sequential, and separate editing paradigms.

Bernini: Latent Semantic Planning for Video Diffusion

Bernini unifies video generation and editing by decoupling high-level semantic planning in the MLLM's native ViT embedding space from pixel rendering in a DiT, augmented by Segment-Aware 3D RoPE and Chain-of-Thought reasoning.

Beyond Distribution Matching: Self-Supervised Representation Forcing for Few-Step Video Generation

Addressing the severe spatial blurring and temporal flickering in few-step video distillation caused by output-only supervision, Self Flow Forcing (SF2) introduces self-supervised representation alignment via asymmetric trajectory noise and a self-bootstrapped EMA teacher without changing the output objective.

Beyond Distribution Matching: Self-Supervised Representation Forcing for Few-Step Video Diffusion

Addressing the spatial degradation and temporal flickering caused by output-only supervision in Distribution Matching Distillation (DMD) under aggressive step compression, this paper introduces Self Flow Forcing (SF2), establishing an information asymmetry via heterogeneous noise levels from a shared trajectory to force the student to predict deep EMA teacher representations, substantially enhancing spatio-temporal coherence with negligible overhead.

Browse all 90 Video Generation papers →


🧩 Multimodal VLM (175)

3D-PLOT-LLM: Part-Level Object Tokens for 3D Large Language Models

3D-PLOT-LLM introduces a decoder-free 3D multimodal large language model that reorganizes point patch tokens into \(K\) spatially coherent regions, inserting reserved vocabulary tokens <part_k> and continuous markers updated by Marker-Space Refinement (MSR) to enable native, bidirectional part-level reading and citation in the LLM's own token stream.

A Closer Look at Dynamic Scene Graph Generation In the Era of Multimodal Large Language Models

This paper revisits Dynamic Scene Graph Generation (DSGG) from task setup and model design, uncovering severe recall-oriented evaluation flaws, introducing 5 practicality metrics, and establishing the first strong top-down reason-then-locate MLLM baseline equipped with Temporal Relation Set (TRS) prediction and Importance-Aware Finetuning (IAF) to achieve SOTA across Action Genome, VidVRD, and PVSG.

A Matched-Budget Audit Framework for Recaptioned Image-Text Supervision Distributions

This paper formalizes recaptioned image-text corpora as supervision distributions induced by captioning policies, introducing a matched-budget five-axis audit framework anchored on Controllable Basic Units (CBUs) alongside an audited ~490M multi-source recaption corpus.

AdaCodec: A Predictive Visual Code for Video MLLMs

AdaCodec introduces an MLLM-oriented predictive visual code interface that allocates full ViT tokens solely to unpredictable reference frames and compact motion-residual P-tokens to intermediate frames, slashing visual token counts by over 84% while cutting latency and outperforming per-frame RGB baselines.

Aligning Forest and Trees in Images: Long Captions for Visually Grounded Understanding

Proposes CAFT, a hierarchical vision-language learning framework that combines a fine-to-coarse superpixel image encoder and a part-whole text chunking encoder to align sub-captions with intermediate visual regions and whole-captions with final scene features, achieving state-of-the-art long-text retrieval and zero-shot referring segmentation without dense region supervision.

AMPS: Adaptive Modality Preference Steering via Functional Entropy

AMPS addresses the fundamental stability-efficiency dilemma of uniform activation steering in multimodal large language models by introducing a functional entropy-derived Modality Contribution Ratio via the logarithmic Sobolev inequality, training a lightweight MLP to adaptively scale steering vectors per instance and achieving robust preference modulation without generation collapse.

Animation2Code: Evaluating Temporal Visual Reasoning in Video-to-Code Generation

Animation2Code introduces the first benchmark comprising 1,069 web animations for evaluating video-to-code de-rendering into executable HTML/CSS/JS, paired with human-aligned metrics that disentangle appearance fidelity from temporal alignment, revealing that frontier VLMs achieve strong appearance similarity (~0.84) but struggle severely with temporal motion reconstruction (~0.31).

Are Multimodal Benchmarks Really Useful? Item-Level Multimodal Benchmark Diagnosis via Structure-Response Co-Calibration

This paper introduces SRCoD, a structure-response co-calibrated diagnosis framework that anchors modality-conditioned Item Response Theory with sample-wise partial information decomposition priors, delivering reliable item-level modality dependence diagnosis and interpretable attribution across multimodal benchmarks.

Are We Making Progress in Multimodal Domain Generalization? A Comprehensive Benchmark Study

Establishes MMDG-Bench, the first unified benchmark covering 6 datasets, 3 task families, 6 modality combinations, and 9 representative methods; training 7,402 models reveals that specialized MMDG methods marginally outperform tuned ERM, trimodal fusion suffers from unresolved competition, and models severely degrade under input corruptions and missing modalities while decoupling accuracy from trustworthiness.

Attention Alignment Between Humans and Vision-Language Models

By implementing a \(2 \times 2\) factorial framework crossing visual encoders (CNN vs. ViT) with language decoders (LSTM with additive bottleneck vs. Transformer with multi-head cross-attention) alongside modern VLMs, this work demonstrates that the top-down decoder attention bottleneck is the primary determinant of human gaze alignment (yielding a 40–50 percentage-point gain), while revealing a striking dissociation between behavioral fixation alignment and cortical neural predictive power.

Browse all 175 Multimodal VLM papers →


🧠 VLM Reasoning (81)

2D Spatial Reasoning with Adaptive Neural Cellular Automata

Addressing the failure of classical Neural Cellular Automata (NCAs) on long-range geometric dependencies caused by rigid local receptive fields, this paper introduces Adaptive Neural Cellular Automata (aNCA), which leverages multi-head modulated deformable convolutions to dynamically discover and reason over spatial constraints with tiny parameter counts, setting new state-of-the-art results on Sudoku and maze navigation.

A Benchmark for Omni-Modal Reasoning in Long Videos

LongShOTBench introduces an omni-modal long-video benchmark covering 274 hour-scale videos (avg. 41 mins), 3,401 intent-driven QA samples across single/multi-turn dialogues with weighted criterion-level rubrics, alongside LongShOTAgent, a training-free search-refine-verify agent that achieves the state-of-the-art score of 66.64% across 105 evaluated systems.

A Systematic Evaluation of Vision-Language Models for Observational Astronomical Reasoning Tasks

Introduced AstroVLBench, a multimodal astronomical reasoning benchmark comprising over 4,100 expert-verified instances across five observational modalities, systematically revealing that frontier VLMs exhibit acute modality dependence, severe bottlenecks in quantitative grounding, and a prevalent "right-answer-wrong-reason" decoupling.

AgroOmni: A Large-Scale Multi-View Agricultural Dataset for Cross-Scale Multimodal Reasoning

To eliminate the severe ground-level bias and scale confusion in agricultural multimodal models, AgroOmni establishes a 288K-pair multi-view VQA dataset bridging Ground, UAV, and Satellite perspectives, enabling the proposed AgroNVILA baseline to achieve a 62.32% SOTA on AgroMind (+15.03% over GPT-5.2).

Allocentric Perceiver: Disentangling Allocentric Reasoning from Egocentric Visual Priors via Frame Instantiation

Allocentric Perceiver is a training-free, plug-and-play framework that leverages off-the-shelf 2D/3D geometric models to recover metric states, dynamically instantiates query-conditioned allocentric reference frames, and prompts VLMs with structured geometric text, successfully offloading mental rotation into explicit computation.

AlloSpatial: Agentic Harness Framework for Spatial Reasoning in Foundation Models

AlloSpatial introduces an agentic harness framework that translates egocentric video observations into queryable Allocentric-Spatial Trees (ASTs) via the World2Mind sandbox, internalizing decoupled evidence collection and cross-modal arbitration through harness-gated reinforcement learning.

ANCHOR: Audio-Visually Grounded Chain-of-Thought Reasoning Benchmark

Introduced ANCHOR, an audio-visual benchmark of 2,319 videos and 8,272 QA pairs integrating SAM3 spatial bounding boxes, temporal intervals, acoustic cues, and human-refined rationales, revealing that models achieving comparable answer correctness frequently exhibit severe perceptual unfaithfulness and reasoning divergence.

ARCANA: A Benchmark for Abstraction and Analogical Reasoning

To evaluate far analogical reasoning between superficially disparate domains, the paper introduces ARCANA, a 200-task benchmark mapping physical principles and real-world objects onto abstract discrete grids, uncovering a significant human-LLM cognitive gap, and proposes the ConceptSIR framework to overcome reasoning local minima via concept slippage, code execution, introspection, and reflection.

ARK: A Dual-Axis Multimodal Retrieval Benchmark along Reasoning and Knowledge

Presents ARK, the first multimodal retrieval benchmark that explicitly decouples knowledge domains from reasoning skills across 17 subtypes and 6 reasoning dimensions, revealing severe bottlenecks of current retrievers in fine-grained visual and spatial reasoning.

ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both

ATLAS internalizes complex visual operations—such as geometric auxiliary lines, region zooming, and symbolic tagging—into single discrete vocabulary tokens without pixel generation or external tool execution, addressing sparse token gradient dilution via Latent-Anchored GRPO (LA-GRPO).

Browse all 81 VLM Reasoning papers →


⚡ VLM Efficiency (11)

Beyond Selection: Token Parameterization for Extreme Visual Token Compression

Rather than selecting a few original patches, Braco constructs a compact visual interface using a DCT low-frequency backbone, basis-coordinate embeddings, budget-dependent coordinate organization, and sparse-pooled spatial residuals; at 576→16 tokens, it retains a Vanilla-normalized aggregate score of 94.0 while reducing full-pipeline prefill latency to 40.59 ms.

Breaking Modality Heterogeneity in Low-bit Quantization for Large Vision-Language Models

To break the modality heterogeneity barrier where vision and text activation outliers reside in disjoint channels and severely distort unified quantization, SplitQ proposes channel-splitting-driven decoupling alongside adaptive cross-modal calibration, maintaining near-lossless accuracy down to W4A4 and W3A3.

EventPrune: Cascaded Event-Assisted Token Pruning for Efficient First-Person Dynamic Scene Understanding

Proposes Event Cascade Pruning (ECP), the first training-free event-assisted pruning framework for first-person dynamic spatial reasoning, which combines causal temporal sampling, motion saliency filtering, and non-parametric rank fusion to outperform the full-token baseline (37.62% vs. 36.31%) at 80% token reduction while delivering a 1.89× inference speedup and 52% GFLOPs savings.

G2TR: Generation-Guided Visual Token Reduction for Separate-Encoder Unified Multimodal Models

G2TR uses VAE latent anchors to select and merge understanding-side visual tokens before LLM prefill in separate-encoder unified multimodal models; retaining 50% of the tokens yields 99.0% relative-average understanding performance and 98.0% editing performance on BAGEL, with approximately 1.94× lower prefill FLOPs, but not lossless performance on every task.

Going Down Memory Lane: Scaling Tokens for Video Stream Understanding with Dynamic Context Memory

Addressing the severe temporal retrieval bias and recall collapse when scaling frame token budgets in streaming video understanding, MemStream introduces training-free adaptive key selection (AKS) for sparse encoding and a reciprocal rank fusion (RRF) mixture-of-experts for frame retrieval, outperforming previous SOTA on CG-Bench and LVBench.

Good Token Hunting: A Hitchhiker's Guide to Token Selection for Visual Geometry Transformers

GoToHunt introduces a training-free hierarchical token selection framework that combines diversity-driven inter-frame selection via farthest point sampling with layer-adaptive intra-frame pruning guided by attention entropy, accelerating visual geometry transformers by over 85% on 500-frame sequences while maintaining or improving reconstruction fidelity.

OmniSelect: Dynamic Modality-Aware Token Compression for Efficient Omni-modal Large Language Models

Addressing the issues of audio-video token explosion and static modality-guided pruning failures in Omni-LLMs, OmniSelect employs a lightweight AudioCLIP to dynamically evaluate query relevance and switch compression regimes, combining fine-grained temporal group pruning to achieve 1.19×–1.33× inference speedup while preserving 94%–99% of full-token accuracy in a training-free manner.

SP2ec: Adaptive Self-Speculative Decoding for Vision-Language Models

Addressing the severe misalignment of small draft models and the suboptimality of fixed draft configurations in multimodal long contexts, this paper introduces SP2ec, a training-free bandit algorithm that exploits the single-peak throughput geometry of self-speculative decoding to adaptively select drafter configurations per query, achieving up to 1.62× lossless inference speedup.

SwiftVLM: Efficient Vision-Language Model Inference via Cross-Layer Token Bypass

SwiftVLM uncovers the non-monotonic layer-wise discriminability and cross-layer importance drift of visual tokens in VLMs, introducing a training-free cross-layer token bypass mechanism with residual offset alignment that temporarily shelves unselected tokens in early layers and re-evaluates them at deeper layers, eliminating localization degradation from premature hard pruning while achieving a ~2× prefill speedup at a 77.8% pruning ratio.

TwinPrune: Density-Aware Two-Phase Token Pruning for Vision-Language Models

To tackle both prefill compute and decode memory bottlenecks in vision-language models, TwinPrune introduces a unified training-free spatial-density signal derived from local cosine dissimilarity that drives density-tilted farthest-point bipartite merging during prefill and single-shot Top-\(k\) KV eviction during decode, establishing a superior compute-memory-accuracy Pareto frontier.

Browse all 11 VLM Efficiency papers →


🎵 Audio & Speech (54)

Adaptive Compression and Targeted Perturbation: A Unified Framework for Generalized Audio Deepfake Detection

To tackle heterogeneous spoofing artifact densities and the pathological robustness-accuracy trade-off in audio deepfake detection, this paper introduces a unified Whisper-based framework integrating an adaptive multi-resolution compression pool (AAFI) and targeted high-frequency mel-band adversarial perturbation (Mel-BAP), achieving state-of-the-art zero-shot cross-domain and cross-lingual detection.

AnyEdit: A Unified Framework for Speech and Singing Voice Editing with Real-World Environmental Consistency

AnyEdit decouples semantic editing from acoustic rendering within a separation-free two-stage framework, combining a noise-robust prosody tokenizer, an autoregressive editor, and an in-context flow-matching renderer to achieve high-fidelity speech and singing editing under real-world noise, reverberation, and background music.

Asymmetric Hierarchical Anchoring for Robust Audio–Visual Cross-Modal Generalization

Addressing the information allocation ambiguity and semantic leakage inherent in symmetric audio-visual frameworks, AHA introduces an asymmetric architecture anchored on the audio RVQ hierarchy, coupled with GRL-based adversarial decoupling and local sliding alignment, to significantly boost zero-shot cross-modal event localization.

Audible World Models: Spatially Aware Sound Generation for 3D Worlds

Audible World Models connects panorama generation, semantic source parsing, dry-audio synthesis, and geometric acoustics to store sound as persistent 3D world state; it produces more accurate motion-dependent directional cues while retaining strong semantic alignment, but currently remains an offline construction system for static worlds.

Audio-Visual Flamingo: Open Audio-Visual Intelligence for Long and Complex Videos

Addressing the short-clip bias and deep-layer visual dominance of existing multimodal LLMs, AV-Flamingo introduces a fully open-source audio-visual foundation model trained on the 7M-instance AV-Skills dataset with a three-stage curriculum and timestamp-grounded Temporal Audio-Visual Interleaved CoT (TAVIT), matching frontier proprietary models across 15+ long-form omnimodal benchmarks.

AudioCALM: Continuous Autoregressive Language Modeling for Universal Audio Generation

AudioCALM establishes the Continuous Autoregressive Language Modeling (CALM) framework by replacing discrete token prediction with a rectified-flow head over continuous audio latents, coupled with block-causal AR-Flow attention and Asymmetric Mixture-of-Modality-Experts (A-MoME) to achieve state-of-the-art unified speech, sound, and music synthesis.

Baton: Explicit Semantic Blueprints for Joint Video-Audio Generation

Baton introduces explicit semantic planning to joint video-audio generation by utilizing an MLLM planner (VA-Planner) to synthesize mutually aligned, keyframe-level perceptual blueprints prior to denoising, combined with Relative Semantic RoPE to bridge asymmetric spatio-temporal grids and prevent cross-modal trajectory drift.

Breaking Adversarial Transferability in Fine-Tuned Speech Recognition

Addressing the vulnerability where fine-tuned private ASR systems are compromised by adversarial perturbations crafted on open foundation models, TransferBreaker combines base-model adversarial fine-tuning, latent Jacobian regularization, and hybrid-gradient adversarial training to slash transfer Word Error Rate (WER) from 92.6% down to 27.8% while preserving clean recognition accuracy.

CS-Dialogue: A 104-Hour Dataset of Spontaneous Mandarin-English Code-Switching Dialogues for Speech Recognition

Introduces CS-Dialogue, the first freely accessible 104-hour large-scale speech dataset of spontaneous Mandarin-English code-switching dialogues with complete transcriptions from 200 speakers, and establishes comprehensive ASR benchmarks across scratch-trained architectures and state-of-the-art foundation models.

CS-Dialogue: A 104-Hour Dataset of Spontaneous Mandarin-English Code-Switching Dialogues for Speech Recognition

To address the scarcity of natural, full-length conversational speech resources in code-switching, this paper presents CS-Dialogue, a 104-hour spontaneous Mandarin-English dialogue dataset featuring 200 speakers with comprehensive transcriptions and multi-paradigm ASR benchmarking.

Browse all 54 Audio & Speech papers →


🔎 AIGC Detection (14)

BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detectors

Introduces the BIAS-ID analysis framework, which systematically quantifies directional transformation biases in AI-generated image detectors by directly analyzing the signed "score shift" of predicted probabilities and the aggregated transform sensitivity (ATS) metric.

BIAS-ID: A Framework for Analyzing Transformation Biases in AI-Generated Image Detection

Addressing the frequent real-world collapse of AI-generated image detectors, this paper introduces BIAS-ID, an evaluation and quantification framework that decouples systematic transformation biases from general robustness decay via score shift and Aggregated Transform Sensitivity (ATS).

BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation

This paper introduces GenBuster-200K, a fair dataset of over 200K high-resolution videos, GenBuster-Bench with progressive dual-axis evaluation, and BusterX, an MLLM baseline trained via entropy-collapse-resistant reinforcement learning (DAPO) that turns long-chain spatiotemporal physical reasoning into the detector itself.

Can Pixels Alone Reveal Image Origin? Minimax Limits and Learnable Interfaces for Passive Provenance

The paper exactly characterizes the optimal robust separation available to pixel-only passive provenance verification through total variation distance, separates this statistical ceiling from interface information exposure in a deployed verifier, and validates the identity in a finite domain without establishing a robustness certificate for realistic images.

FARE: Forensic Acceptance Region Estimation for Catching Bait-and-Switch Image Generators

To tackle deployment-time bait-and-switch risks where black-box image generation APIs covertly replace certified models with cheaper substitutes, FARE constructs a forensic acceptance region from certified outputs alone via constrained residual extraction and boundary-tightening hard positives, achieving near-perfect swap detection under strict 1% false positive rates.

Hitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shift

Addressing the catastrophic failure of supervised AI text detectors under adversarial humanization, new model emergence, and human writing drift, this paper proposes leveraging inference-time homogeneity among unlabeled samples via positive-unlabeled test-time adaptation (TTA), substantially restoring detection and prevalence estimation performance.

LLM-Auction: Generative Auction towards LLM-Native Advertising

LLM-Auction introduces the first learning-based generative auction mechanism that unifies auction allocation and text generation into a single preference alignment framework, internalizing response-level externalities in a single forward pass while achieving state-of-the-art allocation efficiency and provable incentive guarantees.

Log-Likelihood, Simpson’s Paradox, and the Detection of Machine-Generated Text

Uncovers that likelihood-based machine-generated text detectors suffer from Simpson's paradox due to naive token-score aggregation across heterogeneous hidden-space regions, and introduces a lightweight, Bayesian-grounded local calibration framework that boosts Fast-DetectGPT AUROC on GPT-5.4 from 0.63 to 0.85+.

Majority Bit-Aware Watermarking for Large Language Models

To resolve the fundamental trade-off between watermarked text quality and decoding accuracy caused by small green list ratios in perturbation-based multi-bit LLM watermarking, this paper proposes majority bit-aware encoding (MajorMark and MajorMark+), which decouples watermark detectability from green list restriction, guarantees a large green list ratio \(\gamma \ge 0.5\), and significantly improves bit accuracy and text utility.

Post-hoc Selective Classification for Reliable Synthetic Image Detection

To fix the pathological behavior where conventional logit-based confidence scores assign higher confidence to misclassified synthetic images under covariate shifts, ReSIDe extracts layerwise logits via unsupervised linear probing on feature centroids and aggregates multi-layer confidence scores through pairwise preference optimization that provably minimizes an upper bound of the area under the risk-coverage curve (AURC).

Browse all 14 AIGC Detection papers →


🧊 3D Vision (10)

Articraft: An Agentic System for Scalable Articulated 3D Asset Generation

Articraft reframes articulated 3D asset generation into an agentic code-authoring and test-driven repair loop over a dedicated domain-specific SDK, bypassing heavy rendering-based visual inspection to scalably synthesize physics-ready articulated assets, producing Articraft-10K across 245 categories.

BrickAnything: Geometry-Conditioned Buildable Brick Generation with Structure-Aware Tokenization

BrickAnything presents an autoregressive brick generation framework conditioned on 3D point clouds, which reframes global coordinate prediction into local attachment prediction via structure-aware tree tokenization, and integrates dual-objective Direct Preference Optimization (DPO) with parent-aware validity-constrained decoding to produce geometrically faithful and physically stable brick assemblies.

CM-EVS: Sparse Panoramic RGB-D-Pose Data for Complete Scene Coverage

Addressing the severe viewpoint redundancy, depth inconsistency, and unreproducible sampling policies in panoramic 3D visual learning, this paper proposes COVER, a training-free conflict-aware ERP viewpoint curator, and CM-EVS, a panoramic RGB-D-pose dataset that achieves compact, provenance-tracked indoor scene coverage with a median of only ~25 frames under guaranteed submodular approximation bounds.

Direct Conditional Parameterization for N-Dimensional Splatting

Addressing the per-primitive matrix inversion and Schur complement bottleneck in N-dimensional splatting, this work introduces direct conditional parameterization that decouples opacity precision and spatial regression with scale priors, speeding up production slicing by ~7x while ensuring rigorous bijection with joint Gaussian covariances.

GReFEM: Multimodal LLMs as Zero-Shot Semantic Assistants for Physics-Guided 3D Mesh Refinement

GReFEM introduces a zero-shot framework combining multimodal LLMs with physics-guided prompting, trained orthographic view selection, and 2D-to-3D ray projection to steer adaptive 3D volumetric finite element meshing without requiring PDE solves or offline simulation data.

Less Decoder is More Encoder: Geometric Representation Learning from Novel View Synthesis

SNAP identifies that poor geometric representation in novel view synthesis stems from overly expressive decoders and low-level RGB targets, and proposes a pose-conditioned local decoder bottlenecked to a single patch alongside latent feature supervision to learn transferable 3D spatial representations without dense 3D annotations.

Point Cloud Sequence Encoding for Material-conditioned Graph Network Simulators

This paper proposes PEACH, an in-context learning framework that uses a novel 4D spatio-temporal point cloud sequence encoder to extract compact material representations from raw point cloud observations, conditioning a mesh-based graph network simulator to predict full deformable dynamics for unseen materials without any test-time optimization, while demonstrating zero-shot sim-to-real transfer.

RayTun3R: Online Camera Adaptation in 3D Foundation Models

Addressing the severe performance degradation of pretrained 3D foundation models under non-pinhole fisheye optics, RayTun3R reveals an intrinsic pinhole bias in the Jacobian of positional encodings and introduces an ultra-lightweight online camera adapter (~10K parameters) modifying only polar coordinate residuals alongside zero-cost geometric rectifications while keeping the backbone completely frozen.

TrackFish3D: Self-Supervised 3D Tracking of Schooling Fish from Multi-view Videos

TrackFish3D addresses dense multi-camera 3D multi-object tracking of schooling fish with identical appearances and zero annotations by turning calibrated multi-view geometry into self-supervision, achieving 95.8%–96.6% MOTA on SynFish and 81.1% on real zebrafish without any identity supervision or visual appearance backbones.

UnfoldArt: Zero-Shot Recovery of Full Articulated 3D Objects from Text or Image

UnfoldArt resolves articulation ambiguity through hierarchical global-local multi-agent debate and recovers occluded interior geometry by pairing motion-conditioned video generation with 3D latent inpainting, generating interactable articulated URDF assets zero-shot from a single image or text.


🧊 3D Reconstruction & Rendering (58)

3D Point Splatting for mmWave Radar Novel View Synthesis

3D Point Splatting (3DPS) is the first differentiable point renderer for mmWave radar derived directly from the solid-angle radar equation, enabling product-agnostic novel view synthesis across ADC, CRP, and RA signals via closed-form ITU-R P.2040 materials and Hann-windowed PSF splatting in 3 minutes per scene.

Anchor3DGS: Feed-Forward 3D Gaussian Splatting with Compact Anchor-Based Representation

Addressing the primitive redundancy and view-resolution coupling of pixel-aligned feed-forward 3D Gaussian Splatting, Anchor3DGS introduces a compact 3D anchor-based representation guided by local information entropy, occlusion-aware feature aggregation, and lightweight residual refinement, achieving SOTA novel view synthesis with an order of magnitude fewer Gaussians.

Autoregressive Appearance Prediction for 3D Gaussian Avatars

To tackle spurious correlations and abrupt appearance flickering caused by training ambiguities in high-fidelity 3D Gaussian avatars, this paper introduces skinning-weight-based localized pose conditioning and regional PCA, coupled with a convolutional UV texture appearance encoder during training and a causal transformer that autoregressively predicts temporally smooth appearance latents at test driving time.

Birth-Death Structural Learning for 3D Gaussian Splatting

Addressing the disconnect between heuristic density control and empirical reconstruction risk in 3DGS, this paper grounds primitive allocation in the Wasserstein-Fisher-Rao measure variation of the Splat Regression Model (SRM), deriving a first-variation mass score to guide principled birth-death operations and achieve state-of-the-art reconstruction under tight Gaussian budgets.

BLARM: Animating 3D Objects from Video via Blending LAtent Rigid Motion Primitives

BLARM introduces a feed-forward framework that animates static 3D meshes from monocular videos by learning a compact set of time-varying latent rigid motion primitives and time-invariant vertex skinning weights via factorized spatial-temporal attention, achieving rig-free, temporally coherent dynamic mesh reconstruction.

CodecSplat: Ultra-Compact Latent Coding for Feed-Forward 3D Gaussian Splatting

Addressing the inefficiency of compressing irregular 3D Gaussian primitives produced by feed-forward pipelines, CodecSplat integrates neural compression directly inside the generation network by entropy-coding 2D intermediate Gaussian-generation features, achieving controllable high-fidelity novel view synthesis with scene bitstreams of only a few to tens of kilobytes.

CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction

Addressing the severe spatial redundancy and planar hollow artifacts caused by rigid pixel-wise allocation in feedforward 3D Gaussian Splatting, CompactSplat decouples primitive generation into Texture-Aware Partitioning (TAP) and Hierarchical Pyramid Sampling (HPS), tiling flat surfaces with expanded primitives while preserving fine details in textured areas to achieve SOTA multi-resolution fidelity with over 44% fewer primitives and zero-retraining dynamic sparsity control.

ControlFlow3D: Distilling Multi-View Knowledge into Latent Flow Matching for Point Cloud Upsampling

ControlFlow3D formulates point cloud upsampling as conditional flow matching within a frozen 3D latent space, distills multi-view foundation model knowledge into the velocity field via unsupervised shared prototypes during training, and performs linear-complexity coordinate decoding with a Mamba-based architecture for point-only inference.

DinoComplete: 3D Shape Completion with Distilled Semantic Priors and State Space Models

Addressing the challenge that geometry alone cannot reliably infer missing structure in noisy and partial 3D scans, DinoComplete distills 2D foundation model semantic priors into voxel-aligned representations and fuses them via multi-scale voxel state space models, achieving state-of-the-art deterministic shape completion on unseen categories with 108× faster inference than diffusion models.

DySurface: Consistent 4D Surface Reconstruction via Bridging Explicit Gaussians and Implicit Functions

By bridging explicit deformed Gaussians with implicit signed distance functions (SDF) via a dynamic sparse voxel grid (VoxGS-DSDF), DySurface resolves the structural conflict between forward Gaussian motion and backward volumetric rendering, achieving temporally consistent, watertight 4D mesh reconstruction.

Browse all 58 3D Reconstruction & Rendering papers →


🧊 3D Generation & Editing (18)

A Unified Framework for Image-to-3D Part Generation via Variable Granularity

FlexPart introduces a unified image-to-3D part generation framework that supports variable-granularity geometric prompts ranging from sparse points to dense bounding boxes and masks within a single DiT model, enhanced by asymmetric optimal transport guidance to achieve flexible interactive control and high-fidelity part-level generation.

AEON: Unifying Video and 3D World Models

AEON introduces a unified generative world model that maps pretrained video diffusion latents into an explicit 4D Gaussian Splatting latent space via a dual-path domain adapter, generating persistent, photorealistic, and 3D-consistent 4D worlds from text prompts or single images without requiring camera pose inputs.

Controllable Dynamic 3D Shape Generation via 3D Trajectories and Text

T2Mo introduces a feed-forward framework for controllable dynamic 3D shape generation guided by 3D trajectories and text prompts, utilizing a shape-grounded trajectory embedding to eliminate spatial condition leakage and achieve expressive, geometrically faithful mesh animations.

DirectUV: Image-Conditioned UV Texture Generation with Surface-Aware Positional Encoding

DirectUV generates mesh textures directly in the latent UV space of a frozen Flux VAE, embeds 3D surface coordinates into each attention head's rotary positional encoding, and uses the reference image and coarse UV only as conditions, achieving 25.13 PSNR and 0.9527 SSIM on GSO while improving consistency across UV islands and texture detail.

ELSA3D: Elastic Semantic Anchoring for Unified 3D Understanding and Generation

ELSA3D proposes an elastic semantic anchoring framework combining a multiscale octree tokenizer and a lightweight adaptive router, using transient anchor tokens to route decomposed semantic cues into scale-specific geometric evidence and write them back into a unified sequence, achieving state-of-the-art performance across 3D generation and captioning while halving inference compute.

HilbertGen-3D:Hilbert--Multifractal Conditioning for Topology-Aware 3D Generation

HilbertGen-3D introduces a topology-aware 3D point cloud diffusion framework that serializes multi-scale voxel mass measures via 3D Hilbert curves and extracts generalized fractal dimensions as global priors, injecting them into a diffusion Transformer via an efficient Perceiver bottleneck to achieve superior topological coherence and diversity.

Latent Riemannian Flow Matching for Geometry-Grounded 3D Foundation Models

This paper introduces conditional Riemannian Flow Matching directly within the latent manifold of a frozen geometric foundation model (VGGT), demonstrating that respecting its product manifold of four zero-mean hyperspheres eliminates mode collapse and enables high-quality, geometrically consistent 3D scene generation from sparse unposed inputs.

Mesh BDF: Barycentric Dominance Field for 3D Native Mesh Generation

Mesh BDF introduces a continuous scalar field defined on triangular surfaces that bridges discrete topology and continuous diffusion models, enabling high-fidelity native 3D mesh generation with clean topology and unified PBR texture synthesis.

MeshFIM: Local Low-Poly Mesh Editing via Fill-in-the-Middle Autoregressive Generation

MeshFIM introduces the first Fill-in-the-Middle (FIM) autoregressive generation framework for local low-poly mesh editing, leveraging learned boundary markers, topological positional encodings, and a gated geometry difference encoder to achieve seamless boundary attachment and smooth edge flow.

MoCA: Mixture-of-Components Attention for Scalable Compositional 3D Generation

Addressing the quadratic computational bottleneck of global attention in multi-component 3D generation, MoCA introduces importance-based routing to preserve full tokens for Top-k critical components alongside context-preserving compression for distant components, scaling end-to-end generation to 32 fine-grained parts and complex scene instances.

Browse all 18 3D Generation & Editing papers →


🧊 3D Representation & Scene Understanding (27)

CAIRN: Cross-Room 3D Scene Understanding with Topology-Aware Large Multimodal Models

Addressing the limitation of existing 3D-LLMs that flatten multi-room environments into unstructured object tokens, CAIRN introduces a topology-aware 3D-LLM that aligns transformer attention with scene hierarchy using structured masked attention and geometric bias, significantly outperforming prior models on the cross-room benchmark CAIRN-MR.

ClusterSplat: Semantic Cluster Selection for 3D Visual Grounding in Gaussian Splatting

To tackle fragile heuristic thresholding and scattered 3D targeting caused by primitive-level relevance filtering in 3D Gaussian Splatting, ClusterSplat reframes 3D visual grounding as discrete selection over scene-adaptive cluster candidates guided by minimum description length and query-conditioned Transformer scoring, delivering state-of-the-art 3D grounding and rendered 2D segmentation on ScanRefer.

Diffusion Masked Pretraining for Dynamic Point Cloud

DiMP addresses spatio-temporal tube center positional leakage and conditional mean collapse in dynamic point cloud masked pretraining by introducing masked-only center diffusion to prevent positional cheating and formulating inter-frame displacement supervision as a stratified DDPM noise-prediction objective, achieving substantial gains across 4D action segmentation and point-wise semantic segmentation.

DIGS: Distribution-Informed Gaussian Splatting for Training-Free Open-Vocabulary 3D Segmentation

DIGS introduces a training-free open-vocabulary 3D Gaussian segmentation framework that maintains an explicit per-primitive top-M label distribution updated via a tempered Dirichlet rule and renders sharp boundaries with Late-Argmax compositing, constructing reusable instance ontologies in 1.0–3.2 minutes on a single A100 while attaining 71.95% mean mIoU on LERF-OVS.

Emergence of a Shared Canonical Object Frame from In-the-Wild Videos

By routing unannotated in-the-wild object videos through a coarse canonical mesh bottleneck without category-specific details and iteratively solving per-sequence alignments from noisy SfM geometry, a globally shared canonical object frame emerges purely self-supervised without canonical pose annotations or category conditioning.

From Views to Worlds: Active Exploration over 3D Worlds for Vision-Language Models

V2W constructs a VLM-compatible 3D world integrating geometry, poses, and semantics, exposing visual and linguistic interfaces that allow VLMs to actively explore via tool calls and GRPO policy optimization, lifting spatial reasoning from passive 2D views to active 3D world-centric inference.

Geometry Matters in Packed 3D Attention

Addressing the issue where linearizing and memory-packing 3D volumes into token sequences causes attention mechanisms to learn memory storage layout rather than physical object geometry, Pack3D introduces a local-global attention operator decoupling physical coordinates from memory layout via axial grid-coordinate RoPE applied prior to block summarization, outperforming a parameter-matched 3D U-Net on Allen WTC-11 microscopy and achieving a 125x speedup over dense SDPA at 1M tokens.

H2G: Hierarchy-Aware Hyperbolic Grouping for 3D Scenes

H2G introduces a hierarchy-aware hyperbolic affinity field that reconstructs 2D SAM proposals and DINO semantic similarities into tree supervision via Dasgupta's objective, distilling it into a single-stream Lorentz hyperbolic 3D feature field to achieve scale-free, multi-granularity 3D scene grouping.

Instruct-Particulate: Scaling Feed-Forward 3D Object Articulation with Kinematic Control

Addressing training data scarcity and part-granularity ambiguity, Instruct-Particulate introduces kinematic specification conditioning (text prompts, 3D point prompts, and kinematic trees) to scale feed-forward 3D part segmentation and joint parameter estimation across 150k+ heterogeneous assets.

Invaria: Learning Scale and Density Invariance in Point Clouds via Next-Resolution Prediction

Addressing the severe sensitivity of 3D point cloud encoders to sampling density and object scale, Invaria introduces next-resolution latent feature prediction combined with dynamic receptive field calibration, achieving a 56.0% mIoU improvement on ScanNet at 3× lower resolution while cutting model parameters by 45% and input tokens by 40%.

Browse all 27 3D Representation & Scene Understanding papers →


🧊 Depth, Normals & Stereo (7)

Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth

Proposes Depth2Pose, an evaluation framework that measures monocular depth estimation quality via depth-aware relative camera pose solvers without requiring ground-truth depth, and introduces the out-of-distribution D2P dataset featuring challenging statue and vegetation scenes.

DepthMaster: Unified Monocular Depth Estimation for Perspective and Panoramic Images

DepthMaster projects panoramas into tangent patches with a slight 2.5° FoV overlap, injects virtual camera spatial layouts into a standard Transformer, and enforces seamless fusion via a feature-level Correspondence Consistency Loss (CCL), achieving state-of-the-art unified metric depth estimation across perspective and panoramic images without ad-hoc architectural changes.

Focusable Monocular Depth Estimation

This paper introduces Focusable Monocular Depth Estimation (FDE) and the FocusDepth framework, which uses Multi-Scale Spatial-Aligned Fusion (MSSA) to gatedly inject SAM3 box/text prompt features into Depth Anything geometry representations, significantly improving target foreground accuracy and boundary sharpness while preserving global scene geometric coherence.

Learning to Discriminate Scene Structures Makes Self-Supervised Depth Learning Scalable

This paper proposes MaSS, a self-supervised monocular depth estimation framework that bridges a traditional depth encoder with a frozen vision foundation model (VFM) via bidirectional interaction, explicitly decouples contextual and structural representations, and introduces a scene-structure discrimination contrastive objective based on the symmetric KL divergence of softened depth histograms, resolving cross-domain structural bias conflicts and achieving consistent scalability under diverse training data alongside superior zero-shot generalization.

Metric Depth Estimation from Arbitrarily Degraded Low-Resolution Depth Prompts

To tackle arbitrary sensor degradations in low-resolution depth prompts, AdaDS utilizes the contraction property of Gaussian smoothing in forward diffusion, adaptively selecting a starting timestep based on calibration uncertainty and injecting learned noise to project the intermediate latent into high-density regions of a pre-trained diffusion prior for single-step high-fidelity metric depth recovery.

Revisiting Multi-View Stereo: A Sequence-to-Sequence Formulation

This paper reformulates multi-view stereo from conventional reference-view sequence-to-one prediction into a sequence-to-sequence task that jointly outputs geometry for all input views, incorporating ray-map embeddings and a linear-complexity global cost volume to surpass both learning-based MVS and feed-forward baselines.

Selection, Not Fusion: Radar-Modulated State Space Models for Radar-Camera Depth Estimation

By injecting millimeter-wave radar directly into the selective recurrence parameters (step size \(\boldsymbol{\Delta}\) and readout matrix \(\mathbf{C}\)) of Mamba rather than fusing features outside the scan, SemoDepth achieves a 29.9% MAE reduction on nuScenes within 80 m while running at a real-time latency of 26.8 ms.


🧊 Geometric Matching & Localization (6)

BA-T: An Iterative Transformer for Two-View Bundle Adjustment

Drawing inspiration from classical bundle adjustment (BA), BA-T introduces a compact, reusable iterative Transformer layer in implicit token space (using only 38M decoder parameters) that explicitly models latent geometric residuals and alternates between pose and local geometry updates, matching or outperforming deep-stacked feed-forward decoders.

Disambiguating 2D-3D Correspondences in Gaussian Splatting-based Feature Fields for Visual Localization

Addressing the many-to-one 2D-3D matching ambiguity and PnP instability caused by the volumetric extent of Gaussians in photometrically optimized GS feature fields, SplitGS-Loc proposes Mixture-of-Gaussians splitting along the major axis and rasterization weight-based direct feature lifting, building a compact, multi-view consistent feature field in one minute without per-scene optimization to achieve state-of-the-art localization accuracy.

F2G-Pose: Geometry-Aware Foundation Feature Lifting for Direct RGB-D Category-Level Object Pose Estimation

F2G-Pose introduces an end-to-end, single-pass direct regression framework that lifts dense visual foundation features to partial point clouds, models diverse category geometries with a sparse MoE Transformer, and incorporates camera-frame shape completion as dense geometric supervision, achieving real-time 28 FPS inference and state-of-the-art accuracy without iterative diffusion sampling.

GenCOPE: Syn2Real Generalized Category-Level Object Pose Estimation for Robotic Picking

This paper proposes GenCOPE, a synthetic-to-real (Syn2Real) generalized category-level 9DoF object pose estimation framework that enforces training-time 2D and 3D part-level semantic consistency via an LLM and CLIP text encoder to suppress synthetic texture bias, combining a purely global-feature 2D-3D cross-consistency Transformer with an end-to-end decoupled pose decoder to outperform prior art on REAL275 (50.0% mAP at 5°5cm) and Wild6D while running at 40.68 FPS on an onboard 6GB GPU for real-time robotic picking.

MemReg: Streaming Outdoor LiDAR Point Cloud Registration with Hybrid Memory Buffers

To overcome the vulnerability of pair-wise LiDAR registration to dynamic outliers and low overlap, MemReg introduces hybrid memory buffers storing implicit features and explicit poses, unified via Mamba temporal modeling and dynamic history weighting to significantly advance outdoor and indoor registration accuracy.

Register Anything Model for Generalizable and Robust Point Cloud Registration

Addressing the severe domain-shift degradation of traditional point cloud registration models tailored to single environments, Register Anything Model (RAM) establishes a unified cross-domain foundation model using geometry-aware adaptive normalization and an efficient local geometric transformer (EGeoTR), achieving new state-of-the-art generalization across 12 diverse benchmarks spanning indoor, outdoor, and object domains.


🎯 Object Detection (17)

Align3D-AD: Cross-Modal Feature Alignment and Dual-Prompt Learning for Zero-Shot 3D Anomaly Detection

Addressing the severe domain gap between geometric 3D renderings and RGB-pretrained vision encoders, Align3D-AD leverages auxiliary RGB supervision during training via feature alignment with semantic consistency reweighting and couples it with modality-aware dual-prompt contrastive learning, achieving state-of-the-art zero-shot 3D anomaly detection using purely 3D inputs during inference.

Beyond Normal References: Discriminative Few-Shot Anomaly Detection

IDEAL uses a fixed set of few normal and masked anomalous references to suppress local normal variations, extract diverse intrinsic deviation directions, and score their projections, improving seen and unseen anomaly detection without target-domain retraining; under N1A1, it achieves 96.3 / 96.8 image-level / pixel-level AUROC on MVTecAD.

BiHashFormer: Hash-Driven Dual-Branch Transformer for Efficient Object Detection in HRW Shots

To overcome extreme foreground sparsity, computational explosion, and heavy background interleaving in gigapixel-level High-Resolution Wide (HRW) shots, BiHashFormer introduces a hash-driven dual-branch vision transformer that discards half of the key-value pairs in selected RoIs via HashAttention and recovers overlooked targets via HashMiner in non-RoI regions, reducing backbone FLOPs by 37.5% while boosting detection accuracy on the PANDA benchmark.

Contrastive Hypergraph Source-free Domain Adaptive Object Detection in Adverse Weather Conditions

Addressing source-free domain adaptive object detection for autonomous driving under adverse conditions like fog, rain, snow, and night, REUNION formulates high-order contextual modeling via multi-relational hypergraphs and optimizes a group-wise Hyper-InfoNCE objective to propagate semantics from confident instances to corrupted proposals, establishing new state-of-the-art detection accuracy on Cityscapes-Adverse.

COVD: Continual Open-Vocabulary Object Detection with Novel Concept Injection

Addressing the challenge of continually emerging concepts in real-world scenarios without expensive retraining, this paper proposes the Continual OVD with Novel Concept Injection (COVD) task alongside the Novel-114 benchmark, and designs NoIn-Det—a parameter-free expansion framework that freezes the visual encoder and updates only text-branch parameters via stability distillation and knowledge-aware decoupling to assimilate novel concepts while preserving base open-vocabulary capabilities.

Detect Anything in Graphic Design: Element-Level Rewards for Autoregressive Detection

Addressing layer overlap and heavy occlusion in graphic designs, this paper presents DAD, an autoregressive amodal detection model framed as compositional deconstruction, combined with Element Relative Policy Optimization (EleRPO) based on martingale difference decomposition to achieve human-level occluded region recovery and layer decomposition.

Distilling Multi-Teacher Scoring Principles for Unsupervised Time Series Anomaly Detection

Addressing the inference overhead of multi-detector ensembles in unsupervised time-series anomaly detection, AnoMix introduces the CoRank objective to convert normalized multi-teacher score gaps into pairwise preference and consensus confidence supervision, training a compact single-head student (<80k parameters) to directly output scalar anomaly scores that outperform individual detectors and static ensembles across seven benchmarks, with strong zero-shot cross-dataset transferability.

Dynamic Query as Budgeting for Tiny Object Detection

This paper demonstrates that dynamic query mechanisms in tiny object detection primarily function as decoder capacity budgeting rather than fine-grained query selection, and proposes BUTR, an ultra-lightweight, non-learned budgeting-based DETR that outperforms prior complex selection designs with near-zero overhead.

GIFT: Representation Geometry Matters for Single-Domain Generalized Object Detection

Addressing the degradation caused by unconstrained fine-tuning that disrupts pre-trained representation geometry in Vision Foundation Models under single-source supervision, this paper proposes Geometric-Invariance Fine-Tuning (GIFT), which preserves relational geometry via orthogonal Householder reflections and adapts task-specific features via channel-wise scaling, outperforming prior SOTA by +7.9% and +2.9% mPC on Cityscapes-C and DWD with fewer than 1% trainable parameters.

Localize Any Object in X-Ray Security Scans without Human Annotation

To tackle volumetric superposition, weak boundaries, and the severe scarcity of dense annotations in X-ray security inspection, this paper proposes LAO-X, a self-supervised adaptation framework that mines isolated objects from unlabeled scans, synthesizes realistic overlapping scans via the physical Beer–Lambert law, and fine-tunes SAM2 under an occlusion-controlled curriculum with local-to-holistic multi-granularity supervision.

Browse all 17 Object Detection papers →


✂️ Segmentation (25)

COCOTree: A Dataset and Benchmark for Open Tree-Structured Visual Decomposition

This paper introduces the task of open tree decomposition, constructs the large-scale benchmark COCOTree (over 21K images and 1.8M instance nodes) via an automated recursive LVLM-SAM 3 pipeline, and establishes a standardized evaluation metric, Open Tree Quality (OTQ).

Complementing DINO Features with Image Structure for Part Discovery

Addressing the blurry part boundary bottleneck caused by pixel mixing across rigid ViT patch grids, this paper proposes Part-DINO, a training-free framework that complements low-level image contrast-derived geometric regions with DINO semantic features for high-precision part discovery.

Cracks in the Foundation: A Civil Infrastructure Dataset to Challenge Vision Foundation Models

IBM Research and ETH Zurich introduce Cracks in the Foundation (CiF), the largest and most fine-grained civil infrastructure instance segmentation benchmark (≈150k images, 250k masks), revealing that cutting-edge vision foundation models and VLMs suffer from a severe breakdown in zero-shot industrial inspection while specialized supervised models plateau near 25% mAP.

Efficient SAM3 Adaptation for Multi-Class Semantic Segmentation via Dense Competitive Representations

Addressing the prohibitive \(O(N)\) inference latency and semantic misclassifications caused by the absence of inter-class competition in SAM3, this paper proposes DCR-SAM, which decouples visual conditioning to yield shared global visual representations, leverages vision-conditioned text refinement with dynamic pooling, and applies a dense competitive head to achieve a \(33\times\) speedup alongside \(2.3\%\sim 10.7\%\) mIoU improvements.

From Pixels to Concepts: Do Segmentation Models Understand What They Segment?

Introduces CAFE, an attribute-level counterfactual evaluation benchmark that preserves target regions and ground-truth masks while manipulating surface appearance, context, or material composition, revealing that modern open-vocabulary promptable segmentation models consistently rely on visual attribute shortcuts rather than faithfully understanding underlying semantic concepts.

FuseAdapt: Parameter-Efficient Multimodal Fusion for Semantic Segmentation with Missing Modalities

FuseAdapt adapts a frozen vision foundation model for multimodal semantic segmentation by isolating private self-attention low-rank adaptation from receiver-conditioned FFN low-rank residual exchange, complemented by an EMA-supervised Predictive Modality Modeling (PMM) training objective that dramatically improves robustness under missing sensors with zero inference overhead.

Geometry-Guided Semantic Reconstruction for 3D Instance Segmentation

Addressing the issues of cross-modal distribution gaps and passive feature ingestion when directly concatenating lifted 2D foundation model priors, this paper proposes a geometry-guided semantic reconstruction framework that modulates the 2D feature stream using raw 3D geometry and enforces active semantic distillation via masked background superpoint reconstruction.

GMOS: Grounding Moving Object Segmentation in 3D Space and Time

Addressing the heavy reliance of conventional moving object segmentation on pre-computed 2D optical flow/trajectories and coarse sequence-level motion assumptions, GMOS grounds segmentation directly in 3D space and time from raw RGB video, achieving new state-of-the-art results across MOS and fine-grained MOS-I benchmarks while running significantly faster.

Matching-Based Few-Shot Semantic Segmentation Models Are Interpretable by Design

Affinity Explainer (AffEx) exploits the dense cross-attention and correlation matching mechanisms of matching-based few-shot segmentation models alongside multi-level feature ablation to produce pixel-level support attribution maps without backpropagation, establishing new benchmarks on dedicated few-shot causal perturbation metrics.

Pixels to Tokens: Token Space Efficient Active Learning for Low-Budget Semantic Segmentation

TEAL introduces the first ViT-based pixel-level active learning framework for extremely low-budget semantic segmentation by performing two-level MaxHerding diversity selection directly in the native token space to eliminate interpolation artifacts, followed by margin uncertainty refinement on a lightweight decoder grid, surpassing prior baselines by up to +24.36 mIoU with only 1 labeled pixel per image.

Browse all 25 Segmentation papers →


🖼️ Image Restoration (29)

A Principled Self-Referenced Early Stopping Approach for Deep Image Prior

To tackle Deep Image Prior's vulnerability to overfitting measurement noise due to over-parameterization, this paper establishes a self-referenced validation framework that derives three task-tailored early stopping algorithms—channel-similarity reference (CSR-ES), mask reference (MR-ES), and augmented-channel spectral bias reference (ACR-ES)—achieving near-oracle stopping accuracy without requiring ground-truth noise level estimates.

Analyzing and Guiding Zero-Shot Posterior Sampling in Diffusion Models

Under a shift-invariant Gaussian prior assumption, this paper derives closed-form spectral distributions for both zero-shot diffusion posterior samplers and the ideal posterior sampler, replacing ad-hoc hyperparameter heuristics with principled guidance weights optimized by minimizing the averaged Wasserstein-2 distance to the true posterior across image deblurring, inpainting, and accelerated MRI reconstruction.

Beyond Spatial-Domain Supervision: A Relation Constrained Space for Multi-Modal Image Fusion

Rather than treating source images as ideal fused images, RCS-Fusion learns a relational supervision space from frozen DINO/CLIP features and trains different fusion backbones with shared, complementary, and conflict-coordination constraints; its Transformer variant achieves 0.730 mAP in detection using fused M3FD images.

Bricker to BRACE: A Bracket Exposure RAW Dataset and Restoration Model for Flicker-Banding

To tackle flicker-banding artifacts caused by temporal aliasing between rolling shutters and display brightness dimming, this paper presents Bricker—a dual synthetic-real bracketed RAW dataset constructed via physical ray-tracing simulation and automated multi-exposure capture—alongside BRACE, a multi-frame restoration network leveraging frequency-aware priors, multi-scale spatial cross-attention fusion, and a custom evaluation metric SFC.

Can We Model the Artifacts Explicitly? Disentangle Artifacts via Pairwise Edit Residuals

Reinterpreting image manipulation localization as a latent-variable inference problem, this paper introduces the EditGroup-45K dataset with 45K edit groups and proposes Pairwise Artifacts Learning (PAL) with marginal-matching negative sampling to explicitly disentangle manipulation artifacts from semantic content, providing an architecture-agnostic initialization that resists semantic shortcuts.

Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration

Addressing global prompt bottlenecks and discrete MoE routing instability in all-in-one restoration, Continuous Expert Assembly (CEA) synthesizes instance-conditioned low-rank routing bases and residual directions via a lightweight cross-attention hyper-adapter, enabling each spatial token to assemble continuous residual updates through dense signed dot products.

Control-Augmented Autoregressive Diffusion for Data Assimilation

Addressing the compounding error and prohibitive test-time optimization of autoregressive diffusion models in chaotic forecasting, this paper introduces CADA, an amortized stochastic optimal control framework that trains an offline controller to preview future observations and inject affine corrections into diffusion denoising sub-steps, enabling test-time-optimization-free, order-of-magnitude faster data assimilation with superior long-horizon stability.

Exact Posterior Score Estimation for Solving Linear Inverse Problems

This paper derives the closed-form exact posterior score for linear Gaussian inverse problems under general Gaussian interpolants, shows that posterior sampling reduces to anisotropic denoising at an operator-dependent shifted pivot, and proposes EPS, a denoising objective that preserves pretraining structure and achieves superior fidelity and perceptual quality without likelihood gradients or projections.

Flow Map Denoisers: Traversing the Distortion-Perception Plane for Inverse Problems

This paper demonstrates that flow map models implicitly parameterize a one-parameter family of two-time average denoisers via their lookahead parameter \(t\), which provably traces the optimal distortion-perception (DP) frontier for Gaussian targets and enables continuous, zero-shot DP traversal across general inverse problems when embedded into a Plug-and-Play framework.

GLUT: 3D Gaussian Lookup Table for Continuous Color Transformation

By transposing 3D Gaussian Splatting into continuous RGB color space, GLUT models 3D lookup tables via a compact set of explicit learnable 3D Gaussian primitives and local affine mappings, achieving state-of-the-art color fidelity under 1 KB model size, multi-style blending, and training-free real-time local editing.

Browse all 29 Image Restoration papers →


🛰️ Remote Sensing (10)

CoDeRNet: Selective Cross-Task Routing under Heterogeneous Supervision for Change Detection and Captioning

Addressing the asymmetric representation reliability induced by dense pixel-level versus weak sentence-level supervision in remote sensing Change Detection & Captioning (CDC), this paper proposes CoDeRNet, which selectively transfers complementary information via a Confidence-Decomposed Routing (CoDeR) mechanism that balances receiver need and sender reliability, delivering substantial captioning gains while maintaining competitive change localization.

GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation

The first 9-billion-parameter Flow Matching Diffusion Transformer generative foundation model trained entirely from scratch on Earth Observation data, conditioning on continuous geospatial metadata and leveraging a frozen specialist teacher's structural alignment loss to set new state-of-the-art benchmarks in text-to-image synthesis, cloud removal, and SAR-to-optical translation.

How Much of a Model Do We Need? Redundancy and Slimmability in Remote Sensing Foundation Models

This work systematically investigates the post-hoc slimmability of remote sensing foundation models (RS FMs), demonstrating that aggressively reducing Transformer encoder width to 1% of the original compute budget retains over 85% relative downstream performance without any retraining.

LESSViT: Robust Hyperspectral Representation Learning under Spectral Configuration Shift

Addressing the spectral configuration shift across disparate hyperspectral sensors, LESSViT introduces LESS Attention based on structured low-rank decomposition, reducing joint spatial-spectral attention complexity from \(\mathcal{O}(N^2 C^2)\) to \(\mathcal{O}(r N C)\). Paired with channel-agnostic tokenization and physical wavelength-aware RoPE, it retains strong in-distribution capability while markedly enhancing cross-spectral generalization robustness.

LithoBench: Benchmarking Large Multimodal Models for Remote-Sensing Lithology Interpretation

To address the high-barrier demands of domain expertise and causal reasoning in remote-sensing lithology interpretation, this work introduces LithoBench—a 10,000-instance benchmark spanning five cognitive levels—revealing that mainstream vision-language models suffer substantial limitations in higher-order genetic explanation and multi-step geoscientific reasoning.

Manifold Prior Guided Deep Unfolding for Hyperspectral Image Reconstruction

To resolve the scene-agnostic spectral scaling in data-consistency projection and high complexity or limited receptive fields in denoisers of conventional deep unfolding networks, this paper proposes a two-stage diffusion-inferred manifold prior guided deep unfolding framework that injects degradation-free spatial-spectral structural coordinates into both an adaptive spectral-weighting projection step and a linear-complexity Mamba denoiser, achieving state-of-the-art reconstruction on both CASSI and optical filter-based snapshot compressive imaging systems.

PolyTopoBench: A Benchmark for Complex Vector Polygon Generation from Remote Sensing Imagery

Using two aerial-image datasets, four object categories, and 11 public baselines, PolyTopoBench separates region, ring-boundary, and complete ring-structure evaluation, showing that complex instances still have an 84%–100% topology error rate even when their IoU reaches 0.9.

RSRCC: A Remote Sensing Regional Change Comprehension Benchmark Constructed via Retrieval-Augmented Best-of-N Ranking

RSRCC introduces the first remote sensing change question-answering benchmark explicitly designed for localized semantic reasoning with 126k QA instances, constructed via a hierarchical semi-supervised pipeline combining connected component extraction, vision-language screening, and retrieval-augmented Best-of-N ranking.

SatNav: A Scalable Benchmark for Long-Horizon UAV Vision-Language Navigation from Satellite Imagery

SatNav automatically constructs 118,494 long-horizon navigation episodes from satellite imagery and OpenStreetMap and uses SwiftVLN to isolate memory designs: retaining only short-term memory reduces success rate by 21.3 percentage points on both test splits, while real-flight transfer remains supported by just 18 preliminary episodes.

WorldVLN: Autoregressive World Action Model for Aerial Vision-Language Navigation

Addressing dramatic viewpoint shifts, error accumulation, and the lack of causal dynamics in conventional VLAs, WorldVLN introduces a latent autoregressive world action model with a two-stage training scheme and Action-aware GRPO for closed-loop waypoint generation and zero-shot real-drone deployment.


🔍 Anomaly Detection (6)

Feature-Context Consistency: Unsupervised Adversarial Detection with Drift Stability on Attributed Graphs

This paper introduces the Context-Boosted Restricted Boltzmann Machine (CB-RBM) with a gated context pathway, utilizing feature-context consistency as a unified intrinsic signal to achieve high-accuracy, zero-label adversarial detection and drift-stable decision boundaries via closed-form Context Boost (CB) scores.

Flow Mismatching: Unsupervised Anomaly Detection via Velocity Discrepancies in Flow Matching Models

Flow Mismatching bypasses generative reconstruction and memory bank paradigms by treating flow matching as geometric dynamics, directly probing local velocity discrepancies between the learned normal vector field and the geometric path toward a test image to produce robust pixel-level heatmaps and image-level anomaly scores.

GATE-AD: Graph Attention Network Encoding For Few-Shot Industrial Visual Anomaly Detection

GATE-AD establishes a feature reconstruction paradigm for few-shot industrial visual anomaly detection by employing a masked, representation-aligned Graph Attention Network (GAT) to dynamically encode irregular local patch relations, achieving new state-of-the-art accuracy across MVTec AD, VisA, and MPDD while accelerating inference latency by over 25%.

Geometry over Density: Few-Shot Cross-Domain OOD Detection

This paper introduces UFCOD, a framework that replaces density estimation with information-geometric analysis of diffusion trajectories, extracting 2D energy features that measure score function magnitude and temporal smoothness to perform training-free cross-domain OOD detection with ~100 unlabeled ID samples, achieving 93.7% average AUROC across 12 benchmark pairs.

InCLAD: A Continual Learning Benchmark for Industrial Visual Anomaly Detection

To tackle catastrophic forgetting in industrial inspection where inspection systems must extend to new components over time, this paper establishes InCLAD, the first continual visual anomaly detection benchmark spanning 5 datasets, 17 curriculum/random scenarios, and a 27-task cross-dataset stream, alongside native continual adaptations for PatchCore, PaDiM, and CFA.

M3-AD: Reflection-aware Multi-modal, Multi-category, and Multi-dimensional Benchmark and Framework for Industrial Anomaly Detection

M3-AD establishes a unified industrial anomaly detection benchmark across 140 categories with dual reasoning trajectories, and proposes RA-Monitor coupling reflection-aware warm start with reflection-cognitive reinforcement learning (RCRL) to enable models to autonomously correct unreliable initial predictions, substantially improving zero-shot anomaly classification and localization precision.


🧑 Human Understanding (31)

AnyHand: A Large-Scale Synthetic Dataset for RGB(-D) Hand Pose Estimation

This paper introduces AnyHand, a large-scale synthetic RGB-D dataset comprising 2.5M single-hand and 4.1M hand-object interaction images equipped with diffusion pose priors, physics-simulated grasps, full 3D forearm context, and aligned depth maps, which consistently improves state-of-the-art RGB models (HaMeR and WiLoR) on both in-domain and zero-shot cross-domain benchmarks without architectural modifications and sets a new RGB-D SOTA via a lightweight bidirectional depth-fusion module.

AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild

This paper proposes AnyMo, a geometry-aware and setup-agnostic framework for human motion understanding that integrates dense body-surface physics simulation, masked cross-view predictive contrastive pretraining, full-body product-quantized tokenization, and multi-task contrastive instruction tuning, improving average zero-shot Accuracy, macro-F1, and R@2 across 14 unseen human activity recognition benchmarks by 11.7%, 11.6%, and 22.6% respectively, while substantially advancing cross-modal retrieval and wearable motion captioning.

Asymmetric Invertible Threat: Learning Reversible Privacy Defense for Face Recognition

This paper identifies the "asymmetric invertible threat" where existing input-space face privacy cloaks can be neutralized by learned inverse purification or restoration mappings, and proposes the ARFP framework integrating Key-Conditioned Manifold Binding (KMB), Adversarial Restoration-Aware Training (ARAT), and Authorized Reversible Restoration (ARR), achieving a 95.91% mean Protection Success Rate (PSR) across 8 black-box FR models, 91.2% PSR under the strong paired BINN restoration attack, high-fidelity authorized recovery (31.2 dB PSNR, 0.94 SSIM), and nonce-based tamper indication.

AVI-HT: Adaptive Vision-IMU Fusion for 3D Hand Tracking

AVI-HT addresses severe finger occlusion during egocentric hand-object interactions by coupling a two-level hierarchical cross-sensor attention mechanism with kinematic prior masks to fuse 12-channel on-glove 6-DoF IMU measurements with egocentric imagery, decreasing mean keypoint position error by 16.1% on DexGloveHOI.

BiMoGen: Bidirectional Motion-Text Generation via Unified Masked Discrete Diffusion

BiMoGen combines a unified motion-text vocabulary with a bidirectional masked diffusion model for motion generation and captioning, using correspondence learning before conditional generation and generation-aware self-correction to achieve T2M R@1 of 0.555 and M2T CIDEr of 60.2 on HumanML3D, without outperforming existing unified models on every metric.

Breaking the Group Size Barrier: Parameter-Efficient Group Dance Generation with Chain-of-Dancers

ChainDance reformulates music-driven group dance generation as a sequential conditional Chain-of-Dancers, building on a frozen single-dancer diffusion backbone with lightweight role/motion encoders and training-free test-time noise optimization to eliminate group-size rigidity and identity swapping while cutting training time and parameters by 3–6×.

Complexity-guided Regularization for Generalizable Human Gaussian Splatting

To tackle blurry rendering artifacts in high-frequency intricate regions for generalizable human Gaussian splatting, this paper establishes a multi-scale complexity prior from local geometric and textural Shannon entropies of a clothed human mesh, regularizes Gaussian density via pairwise monotonic ranking, and samples complexity-guided local patches for supervision, achieving state-of-the-art novel view synthesis at 34 FPS.

EgoForce: Robust Online Egocentric Motion Reconstruction via Diffusion Forcing

EgoForce addresses the challenge of online full-body motion recovery from sparse, noisy egocentric head and hand tracking by introducing temporally asymmetric noise scheduling inspired by Diffusion Forcing, achieving smooth, drift-free, real-time motion generation under strict causal constraints.

ELMA: Benchmarking Anaphoric Compositions for Long-Term Text-to-Motion Generation

This paper formulates the novel task of long-term text-to-motion generation under anaphoric compositions, introduces the ELMA dataset (13.2k continuous 30-second 3D motion clips with explicit cross-segment dependency annotations) alongside the target-sliced ELMABench benchmark, and demonstrates that fine-tuning existing models with historical text context injection substantially improves cross-segment state persistence and semantic alignment.

EverAnimate: Minute-Scale Human Animation via Latent Flow Restoration

EverAnimate addresses cumulative background degradation and character identity drift in long-horizon pose-guided human animation through Persistent Latent Propagation and Restorative Flow Matching with velocity rescheduling, achieving minute-scale, high-fidelity video generation via lightweight LoRA tuning.

Browse all 31 Human Understanding papers →


📹 Video Understanding (40)

3DSPA: A 3D Semantic Point Autoencoder for Evaluating Video Realism

Addressing the challenge of evaluating generative video realism and physical plausibility without reference videos, this paper introduces 3DSPA—an autoencoder integrating 3D point trajectories, occlusion masks, and DINOv2 visual semantic features to quantify video realism via query reconstruction error (Average Jaccard), significantly outperforming 2D trajectory models and frontier vision-language models in physical rule violation detection and human alignment.

A Principled Optimal Transport Framework for Frame Selection in Long Video Understanding

Addressing the mismatch between long video length and limited visual context budgets as well as the heuristic nature of prior frame selectors, this paper introduces OTFS, a training-free framework that formulates frame selection as semi-unbalanced optimal transport of heterogeneous evidence atoms coupled with coverage-aware exact dynamic programming, substantially outperforming prior strong baselines on Video-MME, LongVideoBench, and MLVU.

Before It Fades: Reinforcing Temporal Representations at Inference Time in VideoLLMs

By quantifying layer-wise representation divergence under frame order reversal, this paper reveals that VideoLLMs acquire temporal information at intermediate layers but allow it to fade toward the output, and proposes Temporal Activation Injection (TAI)—a training-free inference-time intervention that extracts and reinjects temporal vectors along the measured decay schedule to consistently enhance temporal reasoning.

Behavioral Geometric Supervision Aligns Video Foundation Models with Human Social Perception

Introduces Behavioral Geometric Supervision (BGS), a hybrid objective combining local triplet margin constraints and global representational similarity analysis (RSA) via low-rank adaptation (<2% parameters), which aligns video foundation models with human social perception, dramatically outperforming caption-based text embeddings and inducing emergent social-affective dimensions.

Beyond Visual Boundaries: Rethinking Scene Segmentation for Movie RAG

This paper reexamines scene segmentation as the chunking foundation for movie retrieval-augmented generation (Movie RAG), discovers that existing segmentation methods lag behind naive uniform chunking due to over-relying on superficial visual changes, and proposes NarraScene—a narrative-centric benchmark grounded in Event Segmentation Theory that demonstrates substantial gains on downstream reasoning.

CAM: Question Answering on Entity-Centric Videos with Continuous Extraction and Adaptive Querying

Addressing the failure of existing video memory systems to distill long-horizon high-level semantics and recover lost fine-grained details during flat retrieval, CAM introduces evidence-driven asynchronous subgraph abstraction combined with a planner-executor-verifier workflow that adaptively orchestrates topological graph traversal and raw video/audio re-watching, boosting accuracy by up to 23.1 percentage points over SOTA baselines.

CoANeRV: Coordinate-Aware Token-Space Neural Video Representation

CoANeRV introduces a coordinate-aware token-space neural video representation framework that encodes a complete video clip into compact latent tokens via a feed-forward convolutional tokenizer and token former, and reconstructs continuous spatio-temporal queries with a shared coordinate decoder featuring axis-adaptive positional encoding, temperature-modulated cross-attention, and block-wise querying, completely removing per-video optimization while drastically lowering peak attention memory.

Concepts in Motion: Temporal Concept Bottleneck Model for Interpretable Video Classification

Addressing the core dilemma that standard video transformers entangle concept channels and obscure attribution while static concept bottlenecks discard temporal dynamics, MoTIF introduces per-concept diagonal temporal self-attention and VLM-based concept discovery to achieve interpretable and intervenable video classification while bridging the gap to black-box baselines.

Conditional Multi-Event Temporal Grounding in Long-Form Video

To tackle the challenge of localizing and counting all events satisfying complex spatio-temporal conditions in long-form videos, this paper introduces CoMET-Bench along with a cheat-proof Rejection-F1 metric, and presents CoMET-Agent, a training-free search-and-aggregate agentic framework that significantly boosts multi-event temporal grounding precision.

Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization

Addressing the limitations that existing video safety systems are confined to binary classification and lack precision-recall steerability, this paper proposes Adaptive Tversky Policy Optimization (ATPO), introducing feedback-driven Adaptive Tversky Rewards into multi-label GRPO to achieve dramatic performance gains and fine-grained, category-level operating point control.

Browse all 40 Video Understanding papers →


🚗 Autonomous Driving (52)

AutoExpert: Automating 3D LiDAR Annotation from Expert-Crafted Guidelines

AutoExpert turns textual expert rules and a few 2D examples into an adapted detector, then uses a vision-language model to supply instance-specific size and orientation priors for fixed-size multi-hypothesis search, achieving 25.4 mAP3D on AutoExpert-nuScenes without target-task 3D training annotations.

Autonomous Driving Research Requires a Community-Driven Data Paradigm

This position paper systematically analyzes the structural disconnect between citation-monopolizing commercial benchmarks and over 640 unindexed long-tail academic driving datasets, advocating for a community-driven collaborative data ecosystem that decouples raw data collection from task annotation with standardized interoperability and shared infrastructure.

Beyond Chamfer Distance: Granular Order-aware Evaluation Metric For Online Mapping

To address the critical flaws of Chamfer distance (CD) and mAP—namely lack of point-order awareness, susceptibility to outliers, and hard thresholding that masks localization errors—this paper introduces SOSPA, a metric-compliant order-aware single-instance distance, and PLD, a soft multi-instance metric that decomposes into interpretable localization and detection error terms.

Beyond Chamfer Distance: Granular Order-Aware Evaluation Metric for Online HD Mapping

Addressing the lack of order awareness in Chamfer Distance and the loss of positional fidelity caused by hard thresholding in AP, this paper introduces SOSPA, a metric-axiom-compliant ordered polyline distance, and PLD, a multi-instance probabilistic soft-matching metric that decouples mapping error into localization and detection components.

Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving

Addressing the disconnection between at-scale RL driving policies and standardized benchmarks, as well as the simplistic traffic behaviors of current evaluations, this paper introduces BehaviorBench—featuring a nuPlan evaluation interface, an interaction-rich WOMD split (Interactive1k), and heterogeneous traffic agents—uncovering self-play RL generalization failures and establishing a robust hybrid PDM+PPO planner.

BitMTP: When Multi-Token Prediction Meets Low-Bit Large Language Models

Addressing the quality degradation and runtime overheads when naively combining multi-token prediction with low-bit backbones, BitMTP introduces an algorithm–system co-design framework featuring joint NTP-MTP training, flexible decoding strategies, and an accepted-prefix-aligned runtime that delivers 2.1× to 2.4× decode speedups at comparable quality.

Bridging Scene Generation and Planning: Driving with World Model via Unifying Visual Geometry

WorldDrive bridges scene generation and real-time planning by coupling visual and motion representations in a trajectory-aware driving world model, inheriting these mature representations to a downstream planner, and employing a future-aware rewarder to enable millisecond-level foresight evaluation without diffusion denoising.

CausalDriveBench: Evaluating Causal Reasoning in Vision-Language-Action Models for Autonomous Driving

The first benchmark to evaluate causal reasoning in autonomous driving vision-language-action (VLA) models grounded in Pearl's Causal Hierarchy (PCH), featuring causal scene graphs with active/dormant/distractor partitions, 7,285 verified QA pairs, and 1,000 audited counterfactual trajectories that expose severe decoupling between language rationales and motor planning.

Controllable Road Marking Generation

This paper introduces the controllable road marking generation task, combining a latent rectified-flow DiT with topology-aware soft-skeleton supervision and training-free Structured Gaussian Render (SGR) to synthesize high-fidelity lane dividers, road dividers, and pedestrian crossings from drivable geometry, outer-ring context, and text descriptions.

CoWorld-VLA: Thinking in a Multi-Expert World Model for Autonomous Driving

CoWorld-VLA formulates intermediate world reasoning as a multi-expert Latent CoT spanning semantic, geometric, dynamic, and trajectory priors within a VLM, and leverages a two-stream hierarchical diffusion planner to directly condition continuous trajectory generation on these physical world representations.

Browse all 52 Autonomous Driving papers →


🤖 Robotics & Embodied AI (132)

A2Eval: Agentic and Automated Evaluation for Embodied Brain

A2Eval is the first agentic evaluation framework for embodied VLMs that uses a Data Agent to induce an 8-dimensional capability taxonomy and curate a balanced compact suite via diversity-aware clustering, while an Eval Agent synthesizes and validates executable evaluation code, compressing benchmark volume by 85% and speeding up evaluation by 4.6× while correcting leaderboard ranking distortion.

A latent control model for realistic rodent motion

To resolve the trade-off between realism and diversity in embodied animal reinforcement learning, this paper proposes SCAMPER, a dual-prior framework that combines a motion prior for goal-directed control with a behavior prior for exploratory sampling, enabling realistic movement during both goal tracking and sparse-reward exploration.

Action Chunking Proximal Policy Optimization with Feedback Correction

ACPPO-Corr extends PPO with a low-frequency action-chunk planner and a stepwise feedback corrector while retaining a state-only value network, improving final normalized interquartile mean over PPO by 30.4% across 25 simulated robotics tasks—not by 30.4 percentage points of success rate.

Action Images: End-to-End Policy Learning via Multiview Video Generation

Translating 7-DoF robot actions into multi-view pixel-grounded Gaussian action images enables a unified Wan 2.2 video generative backbone to serve directly as a zero-shot manipulation policy without separate action heads.

ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning

ActionUNet constructs a lightweight temporal U-Net in the action feature space to inject hierarchical temporal priors and applies a conditional SIREN continuous decoder with explicit Matérn second-order smoothness constraints, substantially improving VLA robustness against environmental perturbations with minimal overhead.

AdaOcc: Adaptive 3D Occupancy Prediction for Embodied Tasks

AdaOcc proposes an adaptive 3D semantic occupancy prediction framework for embodied tasks that unifies heterogeneous multi-sensor inputs via geometry-guided dual-branch encoding, enables budget-scalable inference through progressive query learning, and enforces high boundary fidelity with a novel containment loss.

AnchorWorld: Embodied Egocentric World Simulation with View-Based Evolution Customization

AnchorWorld integrates hybrid-view projection-based motion control with pose-associated anchor views on top of the Wan2.2 video generation backbone, enabling full-body motion-driven egocentric world simulation with spatially grounded, text-guided local scene state evolution.

APIVOT: Adaptive Planning with Interleaved Vision-Language Thoughts

APIVOT is a VLM-based planner that interleaves language with latent visual thoughts within its autoregressive reasoning trace to internally verify geometric feasibility during planning, significantly outperforming general-purpose VLMs and traditional planning frameworks on complex long-horizon kitchen manipulation tasks.

ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection

Addressing the observation-action modality gap and joint optimization conflicts in predictive vision-language-action models, ATI-VLA introduces a two-stage "Align-Then-Inject" framework that anchors predictive representations into an actionable discrete latent space and adaptively injects them into action decoding under a pure action-centric objective.

BEACON: Cross-Domain Co-Training of Generative Robot Policies via Best-Effort Adaptation

BEACON casts cross-domain robot policy co-training into a discrepancy-aware importance-reweighting problem derived directly from target-domain generalization bounds, significantly improving sim-to-real transfer and out-of-distribution robustness without any explicit feature alignment objective.

Browse all 132 Robotics & Embodied AI papers →


🎮 Reinforcement Learning (237)

A Distribution Mapping Approach to Counterfactually Fair Reinforcement Learning

To eliminate sequential discrimination in high-stakes reinforcement learning without relying on restrictive additive noise assumptions, this paper develops a sequential marginal distribution mapping preprocessing framework (CFSMDM) based on conditional quantile alignment, establishing theoretical finite-sample bounds on suboptimality gap and counterfactual unfairness.

A Geometric Perspective on Reward Function Updates in Inverse Reinforcement Learning

This paper formulates an occupancy-space geometric view of inverse reinforcement learning, proving that moving along the e-geodesic toward the optimal occupancy is equivalent to preconditioning the reward dual gradient with the integral Fisher Information Matrix, which reduces to a computation-free functional convex interpolation under Mirror Descent.

A Single-Sample Polylogarithmic Regret Bound for Nonstationary Online Linear Programming

Addressing nonstationary Online Linear Programming under extreme data scarcity (a single sample per arrival distribution), this paper proposes a dynamic programming-inspired dual re-solving algorithm that leverages a deterministic \(\delta\)-path and uniform dual convergence to establish an \(\mathcal{O}(\log^2 n)\) polylogarithmic regret bound in the large-resource regime.

A Single Deep Preference-Conditioned Policy for Learning Pareto Coverage Sets

This paper establishes the uniqueness and Lipschitz continuity of optimal objective vectors under Smooth Tchebycheff scalarization, deriving an \(\mathcal{O}(1/k)\)-convergent Concave Mirror Descent Policy Iteration (CMDPI) algorithm in occupancy space and its deep actor-critic extensions for dense Pareto-front coverage.

Absolute State-wise Constrained Policy Optimization: High-Probability State-wise Constraints Satisfaction

Addressing the vulnerability of traditional safe RL methods that only enforce constraints in expectation and fail under long-tail violations, this paper proposes Absolute State-wise Constrained Policy Optimization (ASCPO)—the first model-free policy search algorithm that guarantees high-probability state-wise safety constraint satisfaction without assuming known system dynamics, achieved via rigorous non-discounted upper bounds on both the expectation and variance of maximum state-wise costs.

Achieving Optimal Static and Dynamic Regret Simultaneously in Bandits with Deterministic Losses

This paper demonstrates that in oblivious adversarial multi-armed bandits with deterministic losses and known switch count \(S\), one can simultaneously achieve optimal static regret \(\widetilde{O}(\sqrt{AT})\) and optimal dynamic regret \(\widetilde{O}(\sqrt{SAT})\) by leveraging negative static regret and Blackwell approachability.

Active Context Selection Improves Simple Regret in Contextual Bandits

For contextual multi-armed bandits with finite contexts evaluated by context-weighted simple regret, this paper proves that active subpopulation selection yields up to a \(\Theta(k^{1/4})\) gain over passive sampling, derives the optimal power-law allocation \(q_j \propto p_j^{2/3}\) along with a threshold water-filling rule under budget constraints, and develops the EETC algorithm that asymptotically matches the known-distribution fully active rate without prior knowledge of context frequencies.

Actor-Accelerated Policy Dual Averaging for Reinforcement Learning in Continuous Action Spaces

Addressing the computational bottleneck where Policy Dual Averaging (PDA) requires solving a costly optimization sub-problem per continuous action selection, this paper introduces Actor-Accelerated PDA, which approximates sub-problem minimizers via a parameterized policy network and preserves global convergence guarantees while outperforming standard on-policy baselines like PPO.

Advancing Entropy-Level Credit Assignment in RLVR via Proximal Entropy Policy Optimization

AETDICE: Unified Framework and Offline Optimization for Nonlinear Multi-Objective RL

Addressing the long-standing bifurcation between Scalarized Expected Return (SER) and Expected Scalarized Return (ESR) in nonlinear multi-objective RL, this paper introduces the unified Aggregation–Expectation–Transformation (AET) framework and develops AETDICE, the first tractable offline optimization algorithm supporting both trajectory-level nonlinear transformations and expectation-level concave aggregations.

Browse all 237 Reinforcement Learning papers →


🎁 Recommender Systems (24)

Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization

PerTIDE introduces an action-on-item shared event schema and a multi-timescale state-space architecture that routes action-gated evidence into long-, short-, and episodic traces, enabling a single reusable history-update core to serve both predictive ranking and generative headline synthesis with a constant 12 KiB per-user memory cache.

AdaM-Rec: Adaptive Modality Routing for Multimodal Recommendation

AdaM-Rec tackles query-dependent modality reliance in multimodal recommendation by constructing granularity-matched pseudo-queries over historical interactions to dynamically calibrate modality routing weights and retrieval budgets via agentic self-optimization.

An Embarrassingly Simple Graph Heuristic Reveals Shortcut-Solvable Benchmarks for Sequential Recommendation

This paper introduces an intentionally simple, training-free Transition-Graph Heuristic (TGH) that outperforms modern generative and deep sequential recommenders by retrieving few-hop transition neighbors and ranking them with frozen text similarity, uncovering widespread shortcut structures across standard recommendation benchmarks.

Bridging Textual Profiles and Latent User Embeddings for Personalization

This paper introduces BLUE, a reinforcement learning framework that optimizes a generative LLM profiler by uniting embedding-space bidirectional contrastive rewards with text-space next-item prediction signals under GRPO, creating user profiles that are both transparently interpretable and highly effective for retrieval.

Controllable and Content-Based Recommendations

CCBR introduces an editable natural-language user profile bottleneck derived directly from raw multimodal content via MLLMs, plugging into frozen collaborative filtering decoders via tag-supervised pretraining and InfoNCE alignment to deliver competitive accuracy alongside causal, text-driven and multimodal steerability.

DiagnosticIQ: A Benchmark for LLM-Based Industrial Maintenance Action Recommendation from Symbolic Rules

Addressing the industrial IoT bottleneck where fault detection is automated but maintenance action selection requires scarce expert knowledge, this paper introduces DiagnosticIQ—the first benchmark derived from 118 real-world symbolic rules across 16 equipment types comprising 6,690 MCQA items—and exposes through condition inversion that frontier LLMs rely heavily on superficial rule pattern matching rather than genuine logical evaluation.

FoMEMO: Towards Foundation Models for Expensive Multi-objective Optimization

FoMEMO establishes the first foundation model paradigm for expensive multi-objective optimization, pre-training a PFN architecture on hundreds of millions of synthetic GP datasets to predict preference-conditioned aggregated posterior distributions in a single forward pass, enabling zero-shot, update-free in-context optimization.

Goal-Conditioned Supervised Learning for Multi-Objective Recommendation

MOGCSL preserves multiple future session rewards as a goal vector, trains a goal-conditioned next-item predictor with ordinary cross-entropy, and selects inference goals using statistics or two CVAEs, improving purchase prediction and training cost without dominating every click metric or guaranteeing that specified goals are attainable.

Is Your LLM-as-a-Recommender Agent Trustable? LLMs’ Recommendation is Easily Hacked by Biases (Preferences)

This paper introduces BiasRecBench to assess the vulnerability of the LLM-as-a-Recommender paradigm, employing an \(\epsilon\)-Bound quality-margin calibration pipeline to demonstrate that frontier LLMs are heavily susceptible to authority, bandwagon, and instruction biases even when possessing sufficient reasoning ability to identify the ground truth.

LIME: Link-based User-item Interaction Modeling with Decoupled XOR Attention for Efficient Test Time Scaling

LIME proposes a decoupled attention recommendation framework using globally learned "link embeddings" combined with a bidirectional XOR attention mask, reducing user-history self-attention to linear complexity and enabling offline pre-computation of candidate interaction weights, matching heavy Transformer accuracy while delivering a 10× inference speedup.

Browse all 24 Recommender Systems papers →


🔄 Self-Supervised Learning (64)

AmbientFM: A Foundation Model for Ambient Sensing

AmbientFM pre-trains a unified foundation backbone on 9.2 million unlabeled in-the-wild WiFi CSI samples collected over 439 days from 20 commercial IoT device types, combining adaptive frequency aggregation, wireless-tailored contrastive learning, masked reconstruction, and physics-informed autocorrelation prediction to transfer to 9 downstream sensing tasks with >0.90 AUROC under lightweight adaptation.

An Open-Source Training Dataset for Foundation Models for Black-box Optimization

Presents BBO-Pile, the first open-source large-scale pre-training dataset for black-box optimization foundation models (557K trajectories across 3,095 tasks and ~2.5B tokens), and empirically establishes compute and token scaling laws using a family of causal Transformers from 2M to 80M parameters.

Approximation in Contrastive Representation Learning

This paper develops an approximation theory for contrastive representation learning (CRL) by interpreting two-tower inner-product scoring as spectral truncation of a compact pointwise mutual information (PMI) operator, establishing explicit neural network approximation rates and proving fast calibration to downstream retrieval under a density-ratio Tsybakov noise condition.

Baba in Wonderland: Online Self-Supervised Dynamics Discovery for Executable World Models

For symbolic environments where rule vocabularies are prior-misaligned with pretrained semantics, this paper introduces Alice, a closed-loop system that transforms preservation conflicts from rejected program updates into hypothesis classes that simultaneously supply compact counterexamples for code repair and drive class-aware frontier exploration.

Benchmarking Attention for Tabular Foundation Models

Focusing on the unique 2D alternating row and column attention patterns of tabular foundation models, this paper systematically benchmarks 7 attention backends across three GPU generations, quantifies the memory contiguity overhead of strided tensor transpositions, and proposes a decoupled mixed-backend dispatch strategy that yields substantial end-to-end inference speedups.

Beyond IID: How General Are Tabular Foundation Models, Really?

This paper introduces BeyondArena, a unified holistic benchmark comprising 142 curated real-world datasets alongside the DataFoundry framework, systematically probing the true generalization limits of tabular foundation models across IID and non-IID regimes and showing that tuned traditional models still dominate challenging scenarios.

Beyond Isotropy in JEPAs: Hamiltonian Geometry and Symplectic Prediction

Addressing the structural misalignment and predictive unidentifiability caused by enforcing Euclidean isotropic Gaussian marginals in JEPAs, this paper introduces HamJEPA, shifting structural inductive biases from single-view marginals to cross-view Hamiltonian phase-space coupling via symplectic leapfrog integrators with non-isotropic spectral anti-collapse bounds.

Beyond Missing Rates: Rethinking Incomplete Multi-View Clustering with Protocol Divergence

The paper exposes evaluation-protocol divergence in incomplete multi-view clustering through effective missing rate and complete-sample proportion, and introduces CRAFT, which fuses only each sample's observed views: it wins 12 of 13 information-matched incomplete-training conditions, while a separate evaluation tests cross-protocol deployment of checkpoints trained with all views available.

Beyond ViT Tokens: Masked-Diffusion Pretrained Convolutional Pathology Foundation Model for Cell-Level Dense Prediction

Addressing the disruption of spatial continuity and local morphological detail caused by patch tokenization in ViT-based pathology models, this paper proposes CMD, a pure convolutional foundation model using pixel-space masked-diffusion pretraining and frozen pathology semantic guidance that outperforms ViT baselines and end-to-end segmentors under frozen-backbone settings.

Calibrating Scientific Foundation Models with Inference-Time Stochastic Attention

Addressing the failure of deterministic scientific foundation models to provide calibrated uncertainty, this paper introduces Stochastic Attention—an inference-time mechanism that replaces softmax expectations with normalized multinomial sample averages, enabling well-calibrated, sharp prediction intervals on frozen backbones via a 1D post-hoc tuning problem.

Browse all 64 Self-Supervised Learning papers →


📐 Optimization & Theory (313)

A Biconvex Formulation for Stable Transport of Mixture Models with a Unique Solution

Optimal Mixture Transport (OMT) reformulates sample-to-sample optimal transport into a dual-entropy regularized, strictly biconvex optimization problem between subpopulation mixture models, proving a unique global minimizer in a single step with rigorous tilt-stability perturbation bounds and sample-size-independent complexity.

A Differentiable Interior-Point Method in Single Precision

Addressing the severe ill-conditioning and arithmetic breakdowns of standard interior-point methods in low precision, this paper introduces a softplus-based implicit complementarity formulation that produces a spectrally bounded symmetric KKT linear system, enabling robust QP solving and high-fidelity differentiation natively in single precision (float32).

A Kernel Nonconformity Score for Multivariate Conformal Prediction

Introduces the Multivariate Kernel Score (MKS), which lifts Mahalanobis distance to a reproducing kernel Hilbert space (RKHS), proves an equivalence to Gaussian process posterior variance with an exact MMD-KPCA decomposition, and achieves distribution-free finite-sample coverage while drastically compressing prediction region volumes.

A Memory Efficient Unified Algorithm for Online Learning of Linear Dynamical Systems

Addressing high-dimensional linear dynamical systems with a few unstable or resonant modes, this paper introduces a unified online predictor whose memory footprint scales as \(\widetilde{O}(k)\) with the instability complexity \(k\) rather than the full state dimension \(d\), achieving sublinear squared regret without knowing the system matrices.

A Minimal Interpretable Architecture for Zero-Shot Reconstruction of Dynamical Systems

By reverse-engineering the state-of-the-art dynamical systems foundation model DynaMix, this paper derives DynaBase—a minimal, interpretable two-parameter affine nearest-neighbor map that achieves competitive zero-shot attractor reconstruction at negligible computational cost while theoretically reconciling context parroting, flow tracking, and chaotic dynamics within a single mathematical framework.

A Near-optimal SQ Lower Bound for Smoothed Agnostic Learning of Boolean Halfspaces

Under uniform marginals on the Boolean hypercube, this paper establishes a near-optimal Statistical Query (SQ) lower bound of \(n^{\Omega(\log(1+\sigma/\varepsilon^2)/\sigma)}\) for smoothed agnostic learning of halfspaces, matching the \(L^1\) polynomial regression upper bound via an explicit sine-Laguerre dual witness.

A Novel Schur-Decomposition-Based Weight Projection Method for Stable State-Space Neural-Network Architectures

This paper introduces a novel backpropagation-compatible weight projection scheme based on the real Schur decomposition and a pre-factorized formulation for linear discrete-time state-space layers, guaranteeing asymptotic Schur stability with minimal parameter overhead and fast convergence.

A Provably Convergent and Practical Algorithm for Gromov–Wasserstein Optimal Transport

Addressing the gap between exact-projection theory and approximate implementations in nonconvex Gromov–Wasserstein optimal transport, this paper proposes an inexact projected gradient (iPG) framework based on a verifiable primal feasibility residual condition, guaranteeing whole-sequence convergence to stationary points while delivering high sparsity and competitive graph alignment accuracy.

A Reduction from Delayed to Immediate Feedback for Online Convex Optimization with Improved Guarantees

This paper introduces a unified continuous-time event model that decomposes delayed online convex optimization into non-delayed regret and prediction-drift penalties, dramatically improving the delay-dependent regret bound in bandit convex optimization from \(\mathcal{O}(\sqrt{T d_{\max}})\) and \(\mathcal{O}((T d_{\text{tot}})^{1/3})\) down to \(\mathcal{O}(\sqrt{d_{\text{tot}}})\).

A Spectral Framework for Closed-Form Relative Density Estimation

This paper introduces an analytic spectral framework based on generalized eigenvalue decompositions that reformulates the variational estimation of operator-convex f-divergences (including KL) into an integral of weighted chi-squared divergences, yielding closed-form estimators for both divergences and relative log-density potentials purely from first- and second-order feature moments without iterative optimization or partition function evaluations.

Browse all 313 Optimization & Theory papers →


📐 Learning Theory (202)

Sponsored Questions and How to Auction Them

This paper introduces a formal model for auctioning sponsored clarifying questions and subsequent ad slots in conversational AI platforms, proving that an end-to-end VCG mechanism is dominant-strategy truthful and socially efficient, whereas standard modular VCG-per-stage auctions suffer from an unbounded Price of Anarchy.

A Local Geometric Analysis of Maximal Coding Rate Reduction via Error Bounds

This paper proves that the non-convex maximal coding rate reduction (MCR²) objective satisfies a local error-bound property around its local maximizers, guaranteeing that the distance to the non-isolated maximizer set is linearly controlled by the gradient norm, and rigorously establishing local linear convergence for first-order optimization methods.

A Mathematical Theory of Balancing Relational Generalization and Memorization

This paper introduces the transitive inference with exceptions task paradigm, analytically characterizes how kernel ridge regression balances rule generalization and exception memorization across representational geometries, and validates predicted systematic errors in fine-tuned large language models.

A multi-scale information geometry reveals the structure of mutual information in neural populations

By formalizing representational distance contraction under isotropic Gaussian coarse-graining from first principles, this paper derives a unique multi-scale Fisher information metric whose trace rigorously equals population mutual information, and introduces a conditional diffusion estimator that extracts robust, biologically grounded representational geometries from primate visual cortex.

A Theoretical Analysis of Why Masked Diffusion Models Mitigate the Reversal Curse

This paper theoretically establishes that masked diffusion models (MDMs) mitigate the reversal curse through a parameter-level coupling between position-invariant evidence storage and attention-based routing under relative positional encodings, where forward masked gradient steps monotonically decrease the reverse query loss to first order.

A Topological Sorting Criterion for Random Causal Directed Acyclic Graphs

Reveals that popular random DAG generators (Erdős-Rényi and Scale-Free) naturally encode monotonically increasing numbers of "relatives" along the causal order, a topological artifact that leaks causal ordering, collapses Markov equivalence classes (MECs) into singletons, and can be systematically eliminated by time-unrolled DAG benchmarks.

Adaptive inference for functionals of M-estimands

This paper extends inference for smooth functionals of nonparametric M-estimands to adaptive data with known sampling policies, combining one-step bias correction with per-round conditional variance stabilization, or self-normalization when variance converges to a random nonzero limit, to construct asymptotically valid confidence intervals; dynamic pricing simulations show undercoverage for ordinary GLMs and GLMs using inverse propensity weighting alone.

Aligning Inductive Bias for Data-Efficient Generalization in State Space Models

By formalizing linear time-invariant State Space Models (SSMs) as kernel feature maps, this paper proves that their inductive bias is governed by the frequency response and introduces Task-Dependent Initialization (TDI) to align initial filter spectra with task statistics, dramatically improving data-efficient generalization under constrained data budgets.

Ambient Discrete Diffusion: Using the Wrong Data at the Right Time for Data Efficient Learning

Under severe data scarcity, RefineMix introduces out-of-distribution (OOD) data selectively at low diffusion noise levels while keeping high noise levels strictly in-domain, exploiting the fundamental insight that discrete token masking preserves domain separability to eliminate sampling bias and nearly double functional protein yield from only 197 examples.

Amortized Vine Copulas for High-Dimensional Density and Information Estimation

Vine Denoising Copula (VDC) amortizes bivariate edge fitting across high-dimensional regular vines using a single pre-trained U-Net edge denoiser followed by fast Sinkhorn/IPFP marginal projection, turning costly \(O(d^2)\) pair-copula optimizations into millisecond GPU forward passes while preserving exact vine likelihoods and edge-wise mutual information decomposition.

Browse all 202 Learning Theory papers →


🔗 Causal Inference (41)

B-CALM: Bias-Limited Bayesian Borrowing for RCT-Anchored Treatment Effects under Covariate Mismatch

Addressing the challenge where randomized controlled trials (RCTs) lack sample size for conditional average treatment effect (CATE) heterogeneity and observational studies (OS) are confounded with partially overlapping covariates, B-CALM proposes a Bayesian calibrated alignment and borrowing framework that decouples baseline and comparative bias functions, establishes a closed-form information ceiling and saturating effective sample size on observational contrast borrowing, and maintains nominal credible interval coverage without negative transfer.

Beyond the Training Distribution: Evaluating Predictions Under Distribution Shift and Selection Bias

Addressing the compounded challenges of covariate shift and historical decision-induced selective labels in pre-deployment model evaluation, this paper establishes a joint identification framework and derives an efficient influence function to construct an asymptotically normal and semiparametrically efficient double machine learning estimator.

Budgeted Multi-Source Counterfactual Annotation for Off-Policy Evaluation

Addressing limited behavior-policy coverage in contextual-bandit off-policy evaluation, this paper formulates a budget-constrained multi-source counterfactual annotation allocation problem, establishes a first-annotation threshold and four value regimes, and designs a majorization-minimization algorithm that monotonically minimizes estimator variance, reducing evaluation MSE by up to 20.68%.

Causal Abstractions, Categorically Unified

A unified algebraic framework for causal abstractions built upon Markov categories and deterministic natural transformations, rigorously subsuming and generalizing deterministic, probabilistic, non-aligned, and confounded graphical abstractions.

Causal Discovery Under Hard Selection Bias: A New Robust Score-Matching Approach

Addressing observational causal discovery under arbitrary, unknown hard selection bias (truncation), this paper proves that log-density scores and their Hessians remain strictly invariant in the interior of the selection support, and introduces CHess—a robust score-based approach using conditional Hessian rank testing that identifies full causal DAGs across linear and non-linear Additive Noise Models without any selection process knowledge.

Causal Discovery via Transformed Low-Rank Quantile Surfaces

Proposes Low-Rank Quantile Surfaces (LRQS), establishing generic identifiability for bivariate causal discovery under unknown monotone observation distortions and multirank latent shape variations, paired with an alternating truncated SVD and isotonic regression scoring algorithm.

Causal Inference for Sequential Settings under Interference and Latent Confounding

Addressing observational sequential settings plagued by spatial treatment interference, temporal autocorrelation, and low-rank latent confounding, this paper proposes a differentiable framework combining sequential Maximum Pseudo-Likelihood Estimation (MPLE) with low-rank alternating minimization, establishing finite-sample non-asymptotic parameter consistency and generalized treatment effect identification from a single high-dimensional trajectory.

Causal Learning with the Invariance Principle

This paper proves that under acyclicity and bijectivity assumptions, the invariance principle requires only two auxiliary environments (natural experiments) to identify the causal graph and functional mechanisms for arbitrary nonlinear SCMs, guaranteeing identifiability at the counterfactual level.

Causal Multi-Task Demand Learning

To tackle endogeneity in multi-task demand learning where posted prices are systematically correlated with unobserved store-level demand drivers, this paper proposes Decision-Conditioned Masked-Outcome Meta-Learning (DCMOML), which conditions on all endogenous prices while enforcing dual outcome masking and randomized query supervision over locally exogenous price pairs to consistently identify the conditional mean causal parameters.

Causal Representation Learning for Generalisable Recommendation

This paper introduces an information-theoretic causal representation learning objective for recommendation under policy-induced distribution shift, penalising mutual information between item representations and user confounders to discard non-causal leakage and delivering significant online gains on Spotify without any inference-time overhead.

Browse all 41 Causal Inference papers →


🔬 Interpretability (106)

A Hierarchy of Entropy-Shapley Games for Multivariate Predictive Uncertainty

This paper introduces a hierarchy of three entropy-based Shapley games (marginal, sequential, and joint) along with a conditional total correlation game for multivariate probabilistic forecasting, resolving the structural blind spot of component-wise methods that fail to detect feature-driven cross-component dependencies.

ADAPT: Hybrid Prompt Optimization for LLM Feature Visualization

Addressing the severe vulnerability to local minima in discrete prompt optimization for LLM feature visualization, ADAPT introduces a hybrid framework combining multi-group beam search initialization, adaptive gradient-guided and logit-swap mutations, and diversity-preserving slot management, substantially outperforming GCG and EPO across 486 SAE latents in Gemma 2 2B.

Agentic-imodels: Evolving agentic interpretability tools via autoresearch

Agentic-imodels introduces an autonomous research loop where coding agents evolve scikit-learn-compatible tabular regressors optimized for both predictive accuracy and LLM-based simulatability, outperforming classic human-facing interpretable models on the Pareto frontier and boosting downstream agent data science performance on BLADE by up to 73%.

Amortized Linear-time Exact Shapley Value for Product-Kernel Methods

PKeX-Shapley exploits the multiplicative structure of product kernels to define a parameter-free removal operator, computing exact Shapley values for all \(d\) features jointly in \(\mathcal{O}(d^2 n)\) total time (amortized linear time per feature) via prefix-suffix elementary symmetric polynomial recursions, and generalises to MMD and HSIC attribution.

Are Stated Reasoning Steps Causally Load-Bearing? Activation Patching Says Less Often Than Text Editing Does

By introducing targeted interchange activation patching with known counterfactual targets onto self-generated reasoning steps in multi-hop tasks, this work reveals that approximately three-quarters of stated reasoning steps are causally load-bearing, yet standard behavioral text-editing tests overestimate causal faithfulness by 11.4 to 20 percentage points—most severely on easy tasks where reasoning appears most fluent.

Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering

This paper introduces Visual Sparse Steering (VS2), shifting Sparse Autoencoders (SAEs) from passive post-hoc interpretability tools into an actionable downstream intervention space by amplifying active sparse features to steer frozen CLIP embeddings with under 0.1% compute overhead, supported by a closed-form centroid-deviation interpretation and an FVU-based reliability fallback gate.

Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior

Addressing the vulnerability of linear activation steering on curved, anisotropic activation manifolds, INNSteer trains an invertible neural network to map activations into a linearly separable latent space, executing simple translations that pull back into input-dependent nonlinear interventions with exact invertibility and minimal inference overhead.

Beyond Steering Vector: Flow-based Activation Steering for Inference-Time Intervention

By reformulating inference-time intervention from static position-invariant additions into continuous-time velocity field transport conditioned on natural language concepts, FLAS becomes the first learned steering method to consistently outperform in-context prompting on AxBench while uncovering the curved, multi-step, position-dependent geometry of activation space.

Beyond Steering Vector: Flow-based Activation Steering for Inference-time Interventions in LLMs

To overcome the generalization limits of conventional activation steering constrained by single-step, position-invariant, and contrastive assumptions, FLAS introduces a flow-based activation steering framework with a concept-conditioned velocity field integrated over multiple Euler steps, becoming the first learned method to consistently surpass In-Context Prompting on AxBench.

Can Circuit Alignment Predict OOD Generalization?

The paper defines Circuit Alignment Score (CAS) as same-class cross-domain circuit similarity minus cross-class circuit similarity, achieving a mean Spearman correlation of 0.88 on PACS for source-domain model selection without target data, but its OOD ranking consistency requires an additional monotonicity assumption and is not a fully data-free weight diagnostic.

Browse all 106 Interpretability papers →


📦 Model Compression (205)

A Composite Activation Function for Learning Stable Binary Representations

This paper introduces the Heavy-Tailed Activation Function (HTAF), constructed by composing sigmoid and hyperbolic tangent functions, which orthogonally decouples a steep origin gradient for accurate Heaviside approximation from slow tail gradient decay to prevent gradient vanishing, thereby enabling surrogate-free gradient-based training of SNNs, BNNs, and Deep Heaviside Networks while unlocking annotation-free Implicit Concept Bottleneck Models without accuracy loss.

A Free Lunch in LLM Compression: Revisiting Retraining after Pruning

Revisiting the perceived impossibility of post-pruning adaptation in LLMs, this work demonstrates that local reconstruction matching intermediate dense activations matches full PEFT retraining with an order of magnitude less data and compute, revealing a granularity-invariant "free-lunch" regime that makes simple pruning criteria competitive at scale.

A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compression

TACO introduces the first plug-and-play, self-evolving observational context compression framework for CLI agents, dynamically refining structured filtering rules via intra-task implicit feedback and cross-task global pooling to cut token consumption by 12%–27% while boosting task success across six diverse benchmarks without human rules or model tuning.

AC/DC on a Budget – Alternating Sparse Phases

This paper reveals that dense phases in AC/DC sparse training drive critical topology rewiring and parameter sign corrections, and proposes zero initialization, shortened dense phases, and the Alternating Rewiring Compression (ARC) algorithm—which updates only active parameters alongside a small high-gradient fraction of out-of-mask weights—matching or exceeding full-dense AC/DC performance across CIFAR100, ImageNet, and GPT-2 at substantially lower theoretical FLOPs.

Active Learning for Conditional Generative Compressed Sensing

This paper introduces an active learning framework for conditional generative compressed sensing with subsampled Fourier measurements, decoupling prompts into sampling, recovery, and ground-truth roles and proving stable recovery guarantees and explicit sample complexity penalties governed by a prompt compatibility factor.

AdaPreLoRA: Adafactor-Preconditioned Low-Rank Adaptation

Addressing the fundamental obstacle that the rank deficiency of LoRA's generator Jacobian prevents pulling full-weight preconditioners back to factor space, AdaPreLoRA unifies existing optimizers under a constrained affine inversion framework and derives a closed-form factor update minimizing an Adafactor-weighted factor imbalance, achieving the optimal low-rank preconditioned step at \(\mathcal{O}((m+n)r)\) memory.

Adaptive Mass-Segmented KV Compression for Long-Form Reasoning

AMS leaves token-importance scorers unchanged and instead uses attention-derived quality mass to form adaptive segments and allocate retention quotas before local selection, mitigating contiguous context loss in long-form reasoning; on Math500 with a 256-token cache budget, AMS-Expected improves over AdaKV-ExpE2 by 16.0 percentage points.

AdvJudge-Zero: Binary Decision Flips in LLM-as-a-Judge via Adversarial Control Tokens

AdvJudge-Zero leverages a zero-seed beam search over an LLM judge's own conditional next-token distribution to uncover low-perplexity adversarial control tokens that flip binary rejections to approvals, and introduces a 9-class mechanism-stratified LoRA hardening recipe that completely prevents reward hacking collapses under GRPO.

Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation

Reveals that low-bit KV cache quantization in LLMs silently destroys safety alignment while leaving perplexity virtually unchanged, identifies activation channel geometry as the root cause, and introduces the Per-Channel Reduction (PCR) diagnostic framework to prescribe training-free mitigations that recover up to 97% of lost alignment.

Amortized-Precision Quantization for Early-Exit Vision Transformers

Addressing the vulnerability where low-bit quantization noise disrupts early-exit decisions and causes severe accuracy collapse or delayed exits, this paper proposes Amortized-Precision Quantization (APQ) and the MAQEE bi-level optimization framework to jointly optimize exit thresholds and layer-wise bit-widths, slashing BOPs by up to 95% while matching full-precision accuracy.

Browse all 205 Model Compression papers →


🕸️ Graph Learning (44)

Beyond the Golden Teacher: Enhancing Graph Learning Through LLM-GNN Co-Teaching

Addressing the flaw where designating either model as a "golden teacher" propagates its blind spots in few-shot graph learning, this paper proposes LLM-GNN Co-Teaching—a bidirectional framework that dynamically exchanges confident pseudo-labels via architecture-specific small-loss criteria and extracts temporal self-correction preference pairs (RPL-PO) across training rounds, setting a new SOTA across six benchmarks.

Bridging Academia and Industry: A Comprehensive Benchmark for Attributed Graph Clustering

Introduces PyAGC, a production-ready attributed graph clustering benchmark and modular library that unifies algorithms under the Encode-Cluster-Optimize framework, breaks the full-batch memory ceiling via mini-batch neighbor sampling, and evaluates 17 methods across 12 diverse datasets spanning up to 111M nodes.

C_4-Equivariant Flow Matching on Anisotropic Power-Diagram Graphs for Microstructure Generation

This paper models polycrystalline microstructures as compact anisotropic power diagrams (APDs) and learns their joint generative distribution using a \(C_4\)-equivariant graph flow matching network, enabling arbitrary pixel-resolution rendering and training-free gradient guidance to synthesize complex metallurgical morphologies such as welds, cast slabs, 3D-printed layers, and heterogeneous lamellar structures.

Chaining 2-FWL GNNs for Combinatorial Graph Alignment

For the purely structural, unlabelled combinatorial graph alignment problem, this paper introduces a sequential chaining procedure of 2-FWL expressive Folklore GNNs with non-differentiable discrete ranking feedback, breaking through 1-WL expressivity limits on regular graphs and dramatically improving FAQ initialization on hard sparse random graphs.

Chasing Label Shifters: A Change-Aware Framework for Dynamic Graph Node Classification

To address the systematic failure of dynamic graph node classification on "label shifters"—nodes whose class labels transition across snapshots and are overwhelmingly misclassified into their previous classes—this paper identifies neighborhood-aggregation-induced "embedding inertia" as the root cause and proposes CHASE, a plug-and-play, model-agnostic change-aware framework that substantially boosts shifter accuracy while preserving performance on stable nodes.

Concorde: Geometry-Aware Link Prediction via Decoupled Energy Minimization

Addressing the spectral instability and symmetry-breaking dilemma when coupling diffusion operators with geometric representations in sparse graphs, Concorde decomposes link likelihood into three decoupled energy branches—a parameter-free spectral anchor, a mixed-curvature residual, and a tangent geodesic steering field—nearly doubling early retrieval precision on sparse networks.

DiPhon: Diffusion on Graphons for Scalable Graph Generation

DiPhon formulates a continuous diffusion process on graphon space via a Jacobi stochastic differential equation (SDE) and derives a moment-aligned discretized process on finite graphs, enabling graph diffusion models trained on small graphs (20–80 nodes) to generate valid larger graphs up to 300 nodes zero-shot without retraining.

DisRFM: Polar Riemannian Flow Matching for Structure-Preserving Graph Domain Adaptation

Addressing structural degeneration and minimax training instability in Euclidean adversarial graph domain adaptation, DisRFM disentangles radial topological scales from angular class semantics on constant-curvature manifolds and achieves structure-preserving transport via topology-conditioned Riemannian flow matching.

Efficient Dynamic Algorithms for Graph Neural Networks with Non-Linear Propagation

The paper maintains nonlinear propagation fixed points on dynamic undirected graphs using residual certificates, endpoint rescaling, and selective pushes, proves deterministic amortized update bounds under degree-normalized error, and shows that nonlinearity can improve static classification accuracy without substantially increasing dynamic propagation cost.

Efficient Memory Crystallization for Graph Learning under Non-Stationary Distribution Shifts

Addressing the prohibitive runtime and memory footprints of generative replay modules in continual test-time graph adaptation under non-stationary drift, EMC distills incoming graphs into compact, low-rank memory graphs in closed form, leveraging a state-evolving field with provably tighter generalization bounds to achieve SOTA accuracy while slashing runtime by 87.4% and GPU memory by 92.4%.

Browse all 44 Graph Learning papers →


🤝 Federated Learning (9)

Beyond Parameter Aggregation: Semantic Consensus for Federated Fine-Tuning of LLMs

This paper introduces FedCoFiT, a parameter-free federated fine-tuning paradigm that exchanges client text generations on shared public prompts and clusters them in a semantic representation space to form consensus pseudo-labels for local training, decoupling communication volume from model size (yielding up to a 1006× communication reduction on Llama 3.1-405B) while naturally supporting heterogeneous architectures and black-box clients.

BLEND: Balancing Personalization vs. Generalization in Federated Vision–Language Models

BLEND balances client-specific personalization and zero-shot novel-class generalization in federated vision-language models by combining globally aggregated vision/text adapters with a local personalized vision projection head under a dual-objective running anchor loss.

CRAFT: Conflict-Resolved Aggregation for Federated Training

To tackle client gradient conflicts induced by non-IID data distributions, CRAFT reformulates federated aggregation into a reference-anchored constrained least-squares geometric correction problem with a closed-form solution, improving both global model accuracy and tail-client fairness without iterative optimization overhead.

Federate the Router: Learning Language Model Routers with Sparse and Decentralized Evaluations

This paper introduces the first federated learning framework for LLM query routing, enabling decentralized clients to collaboratively train parametric MLP routers and nonparametric K-Means routers using strictly local, sparse, and biased offline query-model evaluation logs, significantly advancing the accuracy-cost Pareto frontier while preserving prompt privacy.

Federated Concept-Based Models: Interpretable models with distributed supervision

Addressing the scarcity of concept supervision and client-side temporal non-stationarity in real-world distributed settings, this paper proposes Federated Concept-based Models (F-CMs), dynamically aggregating causal dependencies, incrementally expanding model architectures, and isolating module updates to enable privacy-preserving, intervenable reasoning over unseen concepts.

Federated Learning by Utility-Constrained Stochastic Aggregation for Improving Robustness

This paper proposes FedUCA, a federated learning framework that formalizes server-side aggregation as a convex feasibility optimization problem to satisfy individual rationality constraints of self-interested clients, exploiting Jensen's gap under Dirichlet-sampled stochastic strategies to induce utility surplus that sustains client participation and avoids federated training collapse under heterogeneity.

HERO: A Heterogeneity-Aware Benchmark Library for Federated Continual Learning

Addressing the lack of fair comparability and entangled confounding factors in Federated Continual Learning (FCL), HERO introduces a benchmark library that orthogonally disentangles task split, client data skew, and task-order mismatch, uncovering pronounced method rank shifts between synchronized and heterogeneous settings.

Ripple in Still Water: Zero-Shot Clustering in Federated Learning with Wavelet Scattering Transform

Ripple introduces an offline clustered federated learning framework that computes client-to-cluster soft assignments entirely prior to training using variance-weighted principal-component prototypes embedded via Wavelet Scattering Transform and decoded by a server-side pre-trained GMM-VAE, matching FedAvg per-round communication cost exactly while enabling zero-shot personalization for unseen clients.

SP-CACW: Convergence-Aware Client Weighting for Selfish Personalized Learning

Addressing negative transfer in non-IID federated learning, SP-CACW establishes a convergence-aware client-weighting framework that directly minimizes an upper bound on the target client's optimization error, dynamically trading off peer gradient bias against stochastic variance via online convex optimization.


📈 Time Series (95)

A Locally Tokenized Generative Model for Robust Time-Series Watermarking

The paper introduces L-VQVAE, a discrete generative model restricting token representations to bounded sliding temporal windows, and LVQMark, a watermarking scheme with alternating parity partitions, unbiased warm-up, and robust re-encoding, eliminating post-editing-induced bidirectional null drift and stabilizing clean-calibrated false-positive detection.

A Scalable Multi-Task Model for Virtual Sensors

Addressing the fragmentation, lack of synergy, and linear scaling costs of conventional virtual sensors, this paper proposes the first unified multi-task model that predicts unseen signals via prototype tokens and learns attention-driven input relevance for structured pruning, delivering up to \(415\times\) speedup and \(951\times\) memory reduction.

A Set-Sequence Model for Time Series

Addressing the challenge of modeling large, high-dimensional cross-sectional time series in finance and engineering, Set-Sequence explicitly decouples permutation-invariant set summarization from per-unit sequence modeling, extracting global cross-sectional latent states in linear time while supporting unaligned series and variable unit sizes at inference without manual feature engineering.

AdaST: Adaptive Coupling for Spatial-Temporal Forecasting

AdaST: Adaptive Coupling for Spatial-Temporal Forecasting

AdaST decomposes spatial-temporal inputs into temporal-specific, spatial-specific, and jointly coupled representations, models them separately, and adaptively recomposes them with correlation-modulated gates, achieving the best reported results on four short-term sensor forecasting benchmarks and reducing PurpleAir MAE from the strongest comparator's 0.511 to 0.489.

AeroChem: Closed-loop Physics-Informed State Space Modeling for Long-term Chemically-Reactive Air Quality Forecasting

Addressing error accumulation, physical drift, and over-smoothing of abrupt pollution spikes in long-horizon chemically-reactive air quality forecasting, AeroChem introduces a dual-branch closed-loop co-correction architecture that recursively fuses probabilistic virtual observations from S-Mamba with UDE physical priors via Neural Kalman Fusion in latent state space.

Arbitrarily Conditioned Hierarchical Flows for Spatiotemporal Events

This paper introduces Arbitrarily Conditioned Hierarchical Flows (ARCH), a hierarchical flow matching framework that decomposes spatiotemporal event generation into temporal and spatial flows paired with hybrid masking, simultaneously achieving arbitrary conditional inference and closed-form, accurate instantaneous event risk (conditional intensity) evaluation.

Benchmarking Sensor-Fault Robustness in Time Series Forecasting

SensorFault-Bench introduces the first cyber-physical system (CPS) sensor-fault robustness benchmark for time series forecasting, featuring 8 standardized scenarios across value, timing, and availability fault families alongside a tripartite metric protocol that decouples relative degradation from absolute error, revealing sharp performance drops in nominal-MSE-favored models and foundational forecasting models.

Beyond Empirical Support: Structured Outlier Generation via Sinkhorn Optimal Transport

SBOG uses Sinkhorn support energy to identify latent boundary anchors and selects locally perturbed candidates with weak support but controlled semantics, raising CARLA's average pointwise F1 across five datasets from 0.362 to 0.432 and improving image out-of-distribution detection; these results measure downstream utility, not recovery of the true outlier distribution.

CAST: Causal Anchored Simplex Transport for Distribution-Valued Time Series

To tackle latent transition-kernel aliasing where identical observed distributions evolve differently across regimes, CAST structures the transition operator directly on the probability simplex via causal successor retrieval, persistence anchoring, and bounded local stochastic transport on ordered supports, substantially curbing multi-step autoregressive rollout drift.

Browse all 95 Time Series papers →


🏥 Medical Imaging (77)

A Steerable Deep Network for Model-Free Diffusion MRI Registration

Presents the first model-free nonrigid diffusion MRI registration framework based on an SE(3)-equivariant steerable UNet, framing registration as an equivariant diffeomorphism over position-and-orientation space with a Fourier-space MMD loss that aligns ensemble average propagators directly from raw data without derived models or explicit reorientation.

Aegis: Generative Gradient Masking for Privacy-Preserving Medical Federated Learning

Aegis adds a task-relevant synthetic-image gradient update after each client-local epoch to increase sample mixing in linear leakage; it reduces the measured reconstruction rate to 9.38%–11.63% across three MedMNIST modalities, without providing zero-leakage or differential privacy guarantees.

Aligning LLMs with Biomedical Knowledge using Balanced Fine-Tuning

Addressing the unique uncertainty structure in biomedical text where dense runs of low-confidence tokens encode crucial entities and mechanistic chains rather than stylistic noise, Balanced Fine-Tuning (BFT) employs local group-normalized token reweighting and sequence-level hard-sample reallocation, delivering superior reasoning and a resilient initialization for sparse-reward RL.

Ask, Answer, and Detect: Role-Playing LLMs for Personality Detection with Question-Conditioned Mixture-of-Experts

ROME introduces an Ask–Answer–Detect framework that leverages LLM role-playing to generate questionnaire-grounded intermediate answers from user posts offline, followed by a question-conditioned Mixture-of-Experts (MoE) network and reliability-importance adaptive modulation to achieve state-of-the-art personality detection with lightweight local inference.

Assessing the robustness of heterogeneous treatment effects in survival analysis under informative censoring

Addressing informative patient dropout in clinical survival analysis, this paper introduces an assumption-lean partial identification framework with sensitivity functions alongside SurvB-learner, a doubly robust and quasi-oracle efficient meta-learner that reliably identifies treatment-responsive subgroups under censoring bias.

Beyond Sparse Captions: Aligning Slide-Level Text and Patch-Level Vision in Pathology

LIMIT (Locked Image Multiple Instance PreTraining) adapts a text encoder to a frozen patch-level visual backbone using report-guided cross-attention pooling and difficulty-weighted InfoNCE, yielding CONCH-Z trained on 335k routine diagnostic reports that sets state-of-the-art zero-shot performance across 22 pathology tasks.

Beyond ViT Tokens: Masked-Diffusion Pretrained Convolutional Pathology Foundation Model

Addressing the issue where Vision Transformer (ViT) patch tokenization disrupts spatial continuity and local cellular morphology, this paper proposes ConvNeXt Masked-Diffusion (CMD), a self-supervised generative pretraining framework using a purely convolutional ConvNeXt-UNet in pixel space conditioned on frozen foundation model semantics, achieving state-of-the-art cell-level dense prediction with minimal task-specific annotation.

B[FM]\(^2\): Brain Foundation Model via Flow Matching with SplitUNet

This paper introduces B[FM]\(^2\), an EEG foundation model that discards conventional discrete patching and masked reconstruction to pretrain directly on raw continuous multi-channel EEG waveforms via optimal-transport conditional flow matching (OT-CFM) and a time-electrode factorized SplitUNet, achieving state-of-the-art performance on 7 of 9 downstream benchmarks with only \(\approx 307\text{ h}\) of pretraining data while generating synthetic EEGs that board-certified neurologists cannot distinguish from real clinical recordings (\(\kappa = -0.096\)).

BigCell: Generating Gigapixel Whole-Slide Images

BigCell introduces the first unified framework capable of natively generating gigapixel whole-slide histopathology images (WSIs), combining a progressively trained 2048-resolution flow-matching Transformer, pathology foundation model embedding-guided sliding-window inference, adaptive CFG velocity caching, and a text-conditioned blueprint generator (CGM).

BrainPro: Process-conditioned Pretraining with Topographic Retrieval for Large EEG Models

Addressing montage heterogeneity and the neglect of experimental protocol context in EEG foundation models, BrainPro couples a universal topographic spatial filter retrieval mechanism with process-conditioned multi-branch masked pretraining, outperforming existing baselines across nine public BCI benchmarks.

Browse all 77 Medical Imaging papers →


🩺 Medical LLM (16)

Ask, Answer, and Detect: Role-Playing LLMs for Personality Detection with Question-Conditioned Mixture-of-Experts

ROME establishes a questionnaire-grounded "Ask, Answer, and Detect" paradigm that leverages role-playing LLMs to simulate offline psychological questionnaire responses and uses a question-conditioned Mixture-of-Experts module to achieve interpretable, highly accurate social media personality detection.

Deceptive Grounding: Entity Attribution Failure in Clinical Retrieval-Augmented Generation

Reveals the deceptive grounding phenomenon in clinical RAG where genuine retrieved clinical trial evidence is misattributed to queried drugs while evading standard faithfulness evaluations, reveals a two-stage permission gate mechanism, and demonstrates entity-attribution verification with 98.7% recall.

Discrete Neural Interlingua for Ontology-Agnostic EHR Modeling

This paper introduces Mediator, a discrete ontology-agnostic EHR tokenizer that maps arbitrary clinical concept text into a shared vector-quantized (VQ) codebook to serve as a neural interlingua, overcoming closed-vocabulary bottlenecks and enabling zero-shot cross-hospital transfer.

DoAtlas-1: A Causal Compilation Paradigm for Clinical AI

Introduces the causal compilation paradigm and the DoAtlas-1 system, which transforms heterogeneous medical evidence from narrative text into executable interventional estimand objects, supporting six typed causal queries with statistical closed-loop validation on the 10,000-participant Human Phenotype Project cohort.

Don’t Pay Attention, PLANT It: Pretraining Attention via Learning-to-Rank

Addressing the failure of random label attention initialization on skewed long-tail classes in Extreme Multi-Label Text Classification, this paper introduces PLANT, a plug-and-play framework that pretrains label attention via Mutual Information Gain and Learning-to-Rank, substantially outperforming larger LLM baselines and prior SOTA models on rare labels.

DT-PBO: an Interpretable Tree-based Surrogate Model for Preferential Bayesian Optimization

Proposes DT-PBO, the first non-linear inherently interpretable tree-based surrogate model for preferential Bayesian optimization, using an innovative pairwise consistency score and straddler exclusion to construct shallow trees with Laplace-approximated zero-sum probabilistic leaves, delivering competitive convergence, up to 400x faster runtime, and transparent clinical triage logic.

EHRNote-ChatQA: A Benchmark for Evidence-Grounded Multi-Turn Clinical Question Answering over Longitudinal Discharge Summaries

EHRNote-ChatQA introduces the first evidence-grounded, multi-turn clinical question answering benchmark spanning patients' longitudinal discharge summaries (967 patients, 16,072 expert-verified QA pairs), showing that LLMs suffer a severe 22.0 pp drop in evidence grounding compared to content answering and exhibit compounding errors across dialogue turns.

Extending ROC Analysis to Uncertainty-Aware Risk Prediction with an Interval-Based AUC (iAUC)

To address the inability of scalar scores to capture epistemic uncertainty in high-stakes clinical risk prediction, this paper extends ROC analysis to interval-valued predictions via the iAUC framework, constructing lower and upper ROC curves that induce a three-region ranking decomposition (confidently correct, ambiguous, and confidently incorrect) and establishing theoretical coverage-based bounds on the Bayes-optimal AUC.

MedEvoEval: Evaluating Continual Evolution of Doctor Agents through Simulated Clinical Episodes

MedEvoEval introduces an executable longitudinal evaluation framework for doctor agents based on action-gated simulated outpatient episodes, revealing process costs hidden by final-answer scoring and demonstrating that accumulated rule memory improves interaction efficiency and composite utility on unseen held-out cases while preserving backward retention.

MedKIT: Evaluating Knowledge Integration and Generalization in Large Language Models

MedKIT converts 6,196 timestamped oncology evidence updates into fact-centered, multi-task probes and compares 12 knowledge integration methods across 5 models, showing that recalling an updated fact usually does not imply using it in open-ended reasoning or generation.

Browse all 16 Medical LLM papers →


🧬 Computational Biology (66)

At FullTilt: Real-Time Open-Set 3D Macromolecule Detection Directly from Tilted 2D Projections

FullTilt feeds aligned 2D tilt-series directly into a visually prompted multiclass 3D detector, replacing volumetric sliding-window detection with cross-tilt row attention, near-zero-tilt query initialization, and training-time geometric augmentation to achieve subsecond zero-shot detection on three real-world cryo-ET datasets, while still requiring simulated-data pretraining and input alignment.

Atom-level Protein Representation Learning Improves Protein Structure Prediction

TriProRep introduces a structure-aware pretraining framework that couples amino-acid identity, backbone geometry, and intra-residue full-atom geometry with generator-corrupted corrective pretraining, delivering consistent gains in homodimer flexible docking, residue-level binding property prediction, and monomer structure prediction distillation on the proposed RepSP benchmark.

Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics

ATMOS is the first all-atom biomolecular dynamics generation framework powered by State Space Models (SSMs), coupling a bidirectional Pairformer latent propagator with conditional diffusion decoding to achieve linear computational complexity and superior fidelity across both protein monomers and protein-ligand complexes.

Benchmarking Compositional Generalisation for Machine Learning Interatomic Potentials

The GMD benchmark systematically evaluates the compositional generalisation of ML interatomic potentials across five controlled tasks—chain extension, fusion, duplication, combination, and fission—revealing that state-of-the-art models and fine-tuned foundation models degrade by 1 to 3 orders of magnitude OOD, with failures driven predominantly by force magnitude extrapolation rather than directional orientation.

Beyond Drug Discovery: The Nanotechnology Molecular Optimization (NMO) Benchmark

NMO tests molecular optimization beyond pharmaceutical priors through three nanophysics tasks constrained by electrode binding, and combines explicit-anchor GGS representations, random-molecule pretraining, and genetic-guided GFNs to discover scientifically promising candidates, although the full model does not maximize AUC on every task.

BioBlobs: Unsupervised Discovery of Functional Substructures for Protein Function Prediction

Branching Flows: Discrete, Continuous, and Manifold Flow Matching with Splits and Deletions

Branching Flows introduces a generative framework based on Generator Matching that transports probability paths over forests of binary trees with learned splitting and deletion rates, enabling variable-length discrete, continuous, and manifold generation while solving the unknown-length infix sampling challenge in antibody and protein design.

Budgeted Quotient-Residual Guidance for Frozen Pocket-Conditioned Molecular Diffusion

Addressing dormant quotient gradients in pocket-conditioned molecular diffusion, Budgeted Quotient-Residual Guidance (QRG) lifts quotient covectors to metric-horizontal ambient directions and caps the correction by the frozen sampler's own step norm, delivering effective inference-time lead-optimization guidance without retraining.

C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation

Inspired by multimodal contrastive learning, C3P aligns bacterial promoter DNA with representations from a frozen pretrained protein language model (ESM2), enabling a compact 100M promoter encoder to substantially outperform 7B generative genome language models in regulatory inference and zero-shot co-regulated gene retrieval.

CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening

Addressing the issue that dual-tower virtual screening easily overfits to global spurious correlations, CausalBind models binder-only observations as a Heckman-style collider V-structure, proves that unconstrained dense interactions are non-identifiable while sparse antichain and low-rank interactions are identifiable, and achieves superior early-enrichment retrieval performance on DUD-E and debiased LIT-PCBA benchmarks via sparse concept-pair bottlenecks.

Browse all 66 Computational Biology papers →


⚛️ Physics & Scientific Computing (61)

APPSolver: Adaptive Patch Partitioning for Point-Wise Ship Flow Prediction on Unstructured Meshes

To alleviate the prohibitive attention cost on large, non-uniform unstructured ship CFD meshes, APPSolver introduces deterministic Adaptive Patch Partitioning (APP) guided by hull distance priors and spatial downsampling, enabling ultra-fast point-wise flow prediction (1.3 ms / 1.8 GFLOPs per forward pass) followed by distance-weighted \(k\)-NN full-reference recovery.

AtomMOF: All-Atom Flow Matching for MOF-Adsorbate Structure Prediction

This paper introduces AtomMOF, the first all-atom variational flow matching model for metal-organic frameworks (MOFs) and host-guest adsorption complexes, which maps 2D molecular graphs of building blocks and adsorbates directly to unconstrained 3D atomic coordinates and lattice parameters and incorporates Feynman-Kac steering guided by machine-learned interatomic potentials (MLIPs) to boost match rate by 35.00% and reduce RMSD by 32.68% on the BW benchmark while outperforming Grand Canonical Monte Carlo (GCMC) by orders of magnitude in sampling speed on ODAC25.

Breakeven Complexity: A New Perspective on Neural Partial Differential Equation Solvers

Addressing the fundamental blind spot where standard accuracy metrics ignore both the substantial up-front training costs of neural PDE surrogates and the low computational cost of low-fidelity numerical solvers, this paper introduces a compute-aware "breakeven complexity" evaluation framework optimized via scaling laws, accompanied by the GPU-native PyFR multi-obstacle benchmark BreakFlow, demonstrating that neural surrogates achieve practical amortization advantages primarily on computationally demanding, high-difficulty physical regimes.

ClothTransformer: Unified Latent-Space Transformers for Scalable Cloth Simulation

ClothTransformer reformulates cloth simulation as autoregressive sequence modeling in a learned latent space, compressing variable-resolution meshes into fixed-size latent tokens via cross-attention and integrating differentiable Continuous Collision Detection (CCD) to deliver unified, scalable, and penetration-free simulation across body-driven garments, robotic manipulation, and free-fall collisions.

CME–SpectrumBench: Can LLMs Analyze Condensed Matter Spectral Data?

Researchers from Stanford and UC Berkeley introduce CME–SpectrumBench, the first expert-designed benchmark evaluating frontier LLMs on research-level ARPES/RIXS spectral data inverse analysis across 190 questions, demonstrating severe degradation under direct dense context reasoning and showing that code-generating agents with Python interpreter access achieve parity with graduate-level handwritten code.

Compute-Optimal Data Scaling for Neural Surrogates via Multi-Fidelity Training

This paper formalizes neural surrogate training as an explicit function of data quantity and fidelity under fixed generation compute budgets, demonstrating that continuous multi-fidelity conditioning yields Pareto-optimal scaling that matches high-fidelity accuracy with up to 6.3× less simulation compute.

Coupled Integral PINN for Discontinuity

Addressing the optimization barrier where strong-form PINNs exhibit diverging residuals and collapse into overly smooth non-physical states at discontinuities, CI-PINN couples a primitive variable network with an auxiliary integral potential network under entropy and adaptive mask constraints, achieving robust mesh-free shock and rarefaction capturing.

Deep-Koopman-KANDy: Dictionary Discovery for Deep-Koopman Operators with Kolmogorov-Arnold Networks for Dynamics

By replacing Deep-Koopman autoencoders with two-layer Kolmogorov-Arnold Networks (KANs) and deriving a chain-rule gradient projection for level-set manifold analysis, this paper decouples representation learning from symbolic description and enables post-hoc dictionary discovery without prior library commitment.

diffusion-warm sampling of the xy model enables fast thermalization at scale

To address the exponential explosion of thermalization times in traditional MCMC for large-scale, low-temperature 2D continuous-spin XY models, this paper proposes using a temperature-conditioned diffusion model trained on small lattices as a warm start, requiring only tens of Wolff MCMC steps to calibrate observables and achieving a tenfold acceleration in thermalization time.

Discovering Phase Space Structure in Learned Hamiltonian Systems

SPS-GAN introduces a conditional generative framework combining port-Hamiltonian dynamics with an extended cyclic-coordinate regularization loss, enabling joint physical trajectory generation and the unsupervised discovery of system degrees of freedom and canonical mechanical coordinates from raw Cartesian observations and video streams without requiring pre-specified phase space dimensions.

Browse all 61 Physics & Scientific Computing papers →


🧮 Scientific Computing (7)

AOT-POT: Adaptive Operator Transformation for Large-Scale PDE Pre-training

Addressing the bottleneck of fitting heterogeneous partial differential equation (PDE) solution operators in unified foundation models, AOT-POT introduces an input-adaptive multi-stream transformation block inspired by classical numerical analysis that projects diverse solution operators into a well-aligned latent space, reducing relative L2 error by 40.9% on average across 12 benchmarks with only 3% parameter overhead.

Conservative Neural Posterior Estimation via Distributionally Robust Training

Addressing the pervasive overconfidence and undercoverage issues in neural posterior estimation under limited simulation budgets, this paper introduces DRO-NPE, a distributionally robust formulation over Wasserstein ambiguity sets that upper-bounds the population risk and KL-based calibration error via a tractable first-order gradient penalty.

Controlling Transient Amplification Improves Long-horizon Rollouts

Addressing the rapid drift and divergence of autoregressive neural physical simulators during long-horizon rollouts, this paper uncovers that the non-normality and non-commutativity of local Jacobians cause transient amplification of perturbations, and proposes zero-inference-cost commutativity regularization to sustain stable rollouts over thousands of steps.

Cross-Fitting for Neural Posterior Estimation

Addressing the extreme overconfidence in Neural Posterior Estimation (NPE) caused by in-sample residual shrinkage during joint mean-variance training, this paper introduces Cross-Fitting NPE (CF-NPE), which decouples conditional-mean regression from out-of-fold residual density estimation via K-fold cross-fitting to resolve the mean underfitting vs. variance overfitting trade-off and achieve calibrated posterior estimation.

Curriculum Multiple Shooting for Robust Training of Neural and Universal Differential Equations

Curriculum Multiple Shooting (CMS) progressively merges overlapping multiple-shooting windows over training stages until recovering the standard single-shooting objective, bridging easy initial optimization on smoothed loss landscapes with full-trajectory fidelity without sensitive continuity penalty tuning.

End-to-End Neural Modeling of EM Response and Design Performance for Free-Form RFIC Passives

To overcome the severe computational bottleneck of electromagnetic simulation in RFIC passive matching networks, this paper introduces an end-to-end framework coupling a frequency-conditioned neural simulator with a circuit-conditioned neural solver, leveraging a complementary response and dependence regularization to break error amplification and deliver tens of thousands of evaluations per second.

MC2: Monte Carlo Correction for Fast Elliptic PDE Solving

The paper introduces MC2, a hybrid WoS-neural network solver that treats low-budget Walk-on-Spheres Monte Carlo solutions as structured noisy estimators and applies a single forward-pass neural correction to eliminate finite-sample variance, matching the accuracy of over 1000x Monte Carlo compute while releasing PDEZoo, a standardized benchmark of 2 million elliptic PDE instances with exact analytic solutions.


🌍 Earth Science (2)

SDECast: Probabilistic Weather Forecasting in Continuous Time with Neural SDEs

Addressing error explosion and broken temporal continuity in discrete autoregressive weather models rolled out at fine time steps, SDECast extends simulation-free SDE Matching to physical state space via concentrated Gaussian boundary anchoring, learning continuous-time atmospheric dynamics without repeated numerical simulation during training and generating calibrated global probabilistic weather ensembles at arbitrary lead times.

ThousandWorlds: A benchmark for climate emulation of potentially habitable exoplanets

ThousandWorlds standardizes 1,689 simulations of tidally locked waterworlds from five global climate models into a low-data parameter-to-mean-climate-field benchmark, measures practical emulator accuracy against same-planet GCM disagreement, and reports a normalized geometric-mean RMSE of 0.462 for GPLFR on the full benchmark.


📡 Signal & Communications (2)

Geometric Inductive Biases for Semi-Supervised Equalization: The Constellation-Aware Transformer

CAT jointly processes the known constellation and received signals from the first layer, with a bidirectional FIR signal branch for inter-symbol interference, reducing pilot requirements through per-block semi-supervised adaptation; the benefit still requires sufficient pilots and receiver-side training, rather than enabling unsupervised or zero-shot decoding across channels.

Learnt Attacks on Quantum Key Distribution under Channel Noise and Device Drift

This paper formulates adaptive eavesdropping on quantum key distribution under continuous channel noise and device drift as a constrained Markov decision process (CMDP), generates a discrete library of compact extremal quantum attack circuits via joint gate-topology and angle search, and uses reinforcement learning to approach the dynamic-programming theoretical upper bound under a zero-detection constraint.


👥 Social Computing (11)

AI-Mediated Communication Can Steer Collective Opinion

This paper demonstrates that popular open-weight LLMs consistently introduce directional biases when editing human texts on contested topics, and proves via an extended Friedkin-Johnsen opinion dynamics model that network propagation amplifies this bias up to 9.2x, systematically shifting long-run collective opinion.

Budgeting Discretion: Theory and Evidence on Street-Level Decision-Making

This paper formalizes street-level discretion as a budgeted finite-horizon dynamic allocation problem, proves that optimal override rates are invariant to welfare scale and determined solely by gain distribution tail shape, and empirically demonstrates using administrative homelessness data that caseworkers dynamically ration overrides in response to short-run capacity openings and administrative timing.

DeEscalWild: A Real-World Benchmark for Automated De-Escalation Training with SLMs

This paper introduces DeEscalWild, the first large-scale benchmark and dataset constructed from in-the-wild police-civilian conflict videos (1,500 full interactions, ~285k dialogue turns), demonstrating that fine-tuned 3B small language models (SLMs) achieve higher behavioral realism and de-escalation efficacy than proprietary LLM baselines like Gemini 2.5 Flash while enabling sub-second edge deployment.

emophone a multi-wave dataset for in-the-wild mobile and wearable affect sensing

EmoPhone introduces a three-wave, in-the-wild multimodal dataset (2020–2022) with 53,139 dense ESM affect labels paired with smartphone and wearable sensing, establishing a three-setting generalization ladder (temporal, within-wave, and cross-wave) that exposes person-level and multi-year distribution shifts.

Enhancing LLMs with Cognitive-Affective Personality Inference for Simulating Human Social-Psychological Behavior

Inspired by the Cognitive-Affective Personality System (CAPS), this paper introduces SPIN, a zero-shot staged inference pipeline that decomposes human participant simulation into task-blind personality compilation, condition-specific cognitive-affective state elicitation, and state-grounded decision readout, substantially improving benchmark-level behavioral alignment across four base LLMs.

Evaluating AI-based Scientific Knowledge Synthesis with Epidemiological Systematic Reviews

Collaborating with the Pathogen Epidemiology Review Group (PERG), this paper introduces AgentSLR, a large-scale benchmark of 16,248 expert-annotated articles on WHO priority pathogens that decouples retrieval, multi-stage screening, and schema-constrained tool-calling extraction to expose structured synthesis bottlenecks across frontier reasoning LLMs.

Freeing the Law with LOCUS: A Local Ordinance Corpus for the United States

This paper introduces LOCUS, the first comprehensive open-source legal corpus covering 9,239 U.S. cities and counties across 7 million pages of local ordinances, constructing a county-harmonized access layer, document-level OCR pipelines, and TrueSkill-based ModernBERT normative regressors to bridge the local law void in legal AI.

Market Incentives for AI Safety Investment

This paper develops a game-theoretic framework of an AI supply chain combining Cournot competition, bilateral Nash bargaining, and neural scaling laws, proving that downstream market competition forces firms to overinvest in safety relative to industry-profit and market-welfare benchmarks, whereas upstream foundation model providers underinvest due to positive externality spillovers.

SLAyiNG: A Diverse and Community-validated Dataset of Queer Slang

This paper introduces SLAyiNG, the first community-validated, ecologically valid dataset of English queer slang covering over 500 terms across 3,405 natural sentences from 20+ subcommunities, empirically proving that low representational bias in language models does not imply linguistic competence on queer vernacular, while showing that lightweight continued pretraining on harmless ingroup slang mitigates both biases simultaneously.

The Bayesian Origin of the Probability Weighting Function in Human Representation of Probabilities

This paper develops a Bayesian encoding-decoding framework on a bounded stimulus domain, demonstrating that the canonical inverse-S probability weighting function in Prospect Theory is the normative outcome of a U-shaped allocation of neural encoding precision (prioritizing probabilities near 0 and 1) integrated via Bayes-risk-minimizing optimal decoding.

Browse all 11 Social Computing papers →


📂 Others (363)

2-Step Agent: How a Bayesian Decision Maker Learns from AI-Decision Support

This paper proposes the 2-Step Agent computational framework for Bayesian causal decision making, derives a closed-form likelihood using auxiliary variables to reduce the exchangeable training plate in multivariate linear-Gaussian settings, and demonstrates that even when the prediction model is well-specified and the decision maker is perfectly rational, a single misaligned prior belief can cause AI decision support to yield worse downstream outcomes than unassisted decision making.

3D-VITAL: Visibility-aware Identity Training from Artificial Liftings of 3D Reconstruction model for Cross-View Object Re-Identification

To tackle geometric corruption and hallucinated texture noise when synthesizing novel viewpoints from single-view 3D reconstructions for Single-View Aerial-Ground ReID (SV AG-ReID), 3D-VITAL extracts a deterministic per-pixel visibility map via ray casting, filtering geometrically broken meshes via Cross-View Consistency Filtering (CVCF) and guiding ViT identity attention to observed regions via Visibility-Aware Attention Supervision (VAAS) to establish new state-of-the-art accuracy on AG-ReID, AG-ReID.v2, and MOO.

A Scalable Nonparametric Continuous-Time Survival Model through Numerical Quadrature

QSurv directly parameterizes the continuous-time instantaneous hazard and computes the cumulative hazard via Gauss-Legendre numerical quadrature, coupled with Time-LoRA to eliminate redundant backbone evaluations for scalable, end-to-end survival modeling on high-dimensional clinical data.

A simple model of co-emergence of grid and place fields

Presents the first recurrent network respecting Dale's Law that achieves the co-emergence of entorhinal grid cells and hippocampal place cells under a single next-sensory prediction objective without spatial supervision or pre-existing cell priors.

A Stratified Multi-Rater Evaluation of LLM-Based Virtual Standardized Patients with a Deployed Data-Generation Platform

To overcome the lack of expert baselines and fine-grained failure attribution in virtual standardized patient (VSP) evaluations, this paper establishes a double-blind, fully crossed multi-rater framework with 17 senior medical students (7,480 ratings) and an open-source clinical practice platform, uncovering pronounced LLM deficiencies in clinical reasoning, patient counseling, and negative question polarity.

A Systematic Evaluation of Co-folding Model Representations for Small-Molecule Learning

This paper systematically evaluates and transfers atom-level ligand representations from Boltz2, a modern protein-ligand co-folding foundation model, to standalone small-molecule tasks, demonstrating state-of-the-art ADMET property prediction, \(2\times\) faster molecular diffusion training, and improved sample efficiency in structure-guided reinforcement learning.

A Unified Uncertainty Representation for Graph Neural Networks via Doubly-Spectral Stochastic Expansion

DSS-GNN expands node representations along graph-frequency and stochastic chaos-order axes and propagates uncertainty through deterministic quadrature; its standalone mode obtains the lowest mean Brier among compared methods on 14 calibration benchmarks, while its hybrid mode outperforms the listed baselines on most node-OOD settings and all 7 GOOD concept-shift settings.

Active Learning for Gaussian Process Regression Under Self-Induced Boltzmann Weights

This paper formalizes active learning under self-induced Boltzmann distributions, deriving a closed-form acquisition function (AB-SID-iVAR) that avoids partition function estimation and guarantees sublinear convergence in both discrete and continuous domains.

Active Learning from Positive and Unlabeled Examples

Addressing weakly supervised binary classification with one-sided probabilistic feedback (positive and unlabeled data), this paper presents the first theoretical framework and algorithms for active PU learning, establishing rigorous PAC label complexity bounds that incur only an additional factor proportional to the disagreement coefficient \(\theta\).

Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery

To tackle geospatial target discovery under tight sampling budgets, non-revisitable streaming inputs, and distribution shift, this paper formalizes the OWL-GPS setting and introduces a CVAE-based concept relevance encoder coupled with relevance-guided active sampling and semantic-clustering meta-batch formation for rapid online adaptation.

Browse all 363 Others papers →