CVPR2025 CVPR2025 accepted papers CVPR2025 paper list AI paper notes top conference papers Image Generation Multimodal VLM 3D Reconstruction & Rendering Human Understanding Segmentation Autonomous Driving Video Generation Medical Imaging

📷 CVPR2025 Accepted Papers¶

1819 CVPR2025 paper notes covering Image Generation (313), Multimodal VLM (137), 3D Reconstruction & Rendering (121), Human Understanding (109), Segmentation (96), Autonomous Driving (93), Video Generation (89), Medical Imaging (82) and other 49 areas. Each note has TL;DR, motivation, method, experiments, highlights, and limitations — 5-minute reads of core ideas.

💡 LLM Reasoning (7)¶

Argus: Vision-Centric Reasoning with Grounded Chain-of-Thought: Argus proposes a grounded visual CoT mechanism that enables explicit target-oriented visual attention by first making the MLLM predict a question-related bounding box (RoI), and then resampling/re-encoding the visual tokens of that region as reasoning context, achieving dual SOTA in visual reasoning and object grounding among 7B/8B-scale MLLMs.
Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation: AoTD uses an LLM agent to decompose complex video questions into subtasks, invokes expert vision models to execute them, and collects intermediate results as a Chain-of-Thought (CoT). After quality filtering using an LLM, the CoT is distilled into a Video-LLM, enabling the end-to-end model to achieve both accurate answers and interpretable multi-step reasoning capabilities.
Interleaved-Modal Chain-of-Thought: Proposes Interleaved-Modal Chain-of-Thought (ICoT), which interleaves image region crops as visual rationales within reasoning steps. By using a parameter-free Attention-driven Selection (ADS) to intelligently select and insert key regions from the input image into the generated sequence, it achieves up to a 14% improvement over existing multimodal CoTs on Chameleon and Qwen2-VL.
Reason-before-Retrieve: One-Stage Reflective Chain-of-Thoughts for Training-Free Zero-Shot Composed Image Retrieval: This paper proposes OSrCIR, a training-free one-stage zero-shot composed image retrieval method. It utilizes multimodal large language models to directly process the reference image and modification text, and accurately understands the user's implicit intent through reflective Chain-of-Thought reasoning, outperforming existing training-free methods by 1.80% to 6.44% across multiple benchmarks.
Style Evolving along Chain-of-Thought for Unknown-Domain Object Detection: This paper proposes a Chain-of-Thought Guided Style Evolution (CGSE) method. By generating three-level progressive style descriptions (word $\rightarrow$ phrase $\rightarrow$ sentence), combined with feature disentanglement and class-specific prototype clustering, CGSE achieves significant performance improvements in domain generalization object detection on five adverse weather scenarios and the Real-to-Art benchmark.
VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection: VideoEspresso constructs a large-scale video CoT reasoning dataset of over 200k samples (containing spatial bounding box and temporal grounding annotations). It also proposes a hybrid framework, VideoQA-SC, which employs a lightweight 1.5B model to select an average of 2.36 core frames, followed by an 8B reasoning model performing two-stage evidence extraction and answer generation. With only 1.8% of the frames and 14.7% of the computation, it outperforms GPT-4o and all open-source LVLMs.
Learning-enabled Polynomial Lyapunov Function Synthesis via High-Accuracy Counterexample-Guided Framework: This paper proposes a learning-enabled polynomial Lyapunov function synthesis method which combines learning and verification. It uses data-driven machine learning to guide the selection of polynomial forms and iteratively optimizes them through a high-accuracy counterexample-guided framework, striking a balance between flexibility and mathematical rigor.

🦾 LLM Agent (12)¶

GUI-Xplore: Empowering Generalizable GUI Agents with One Exploration: Proposed the GUI-Xplore dataset (312 applications, 32K+ QA pairs, 5-level tasks) and the Xplore-Agent framework (Action-aware GUI Modeling + GUI Transition Graph Reasoning). By simulating the human strategy of "exploring before reasoning", it improves StepSR by approximately 10% on unfamiliar applications compared to state-of-the-art GUI agents.
ATA: Adaptive Transformation Agent for Text-Guided Subject-Position Variable Background Generation: Proposes the ATA (Adaptive Transformation Agent) framework to achieve precise control over subject position and pose in text-guided background generation, dynamically adjusting the subject's placement in the background via an adaptive transformation module while balancing visual consistency and semantic plausibility.
ComfyBench: Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems: ComfyBench proposes the first comprehensive benchmark (200 tasks, 3205 node documentations, and 20 curriculum workflows) to evaluate the capability of LLM-based agents to autonomously design collaborative AI systems in ComfyUI. It also introduces the ComfyAgent framework, which leverages code-based workflow representation and multi-agent collaboration to achieve a resolve rate comparable to o1-preview. However, it resolves only 15% of helper creative tasks, highlighting a significant gap in autonomous system design for LLM agents.
Feature4X: Bridging Any Monocular Video to 4D Agentic AI with Versatile Gaussian Feature Fields: This paper proposes Feature4X, a versatile framework that distills the functionalities of various 2D visual foundation models (e.g., SAM2, InternVideo2) from arbitrary monocular videos into a unified 4D Gaussian feature field via a dynamic optimization strategy. This work represents the first attempt to lift video foundation models to 4D features based on Gaussian Splatting, supporting segment anything from novel views, geometric/appearance editing, and free-form VQA.
NADER: Neural Architecture Design via Multi-Agent Collaboration: NADER models neural architecture design as a multi-LLM-agent collaborative task: a Reader extracts knowledge from papers, a Proposer generates improvement plans, a Modifier implements modifications using Directed Acyclic Graphs (DAGs), and a Reflector learns from failures. With only 10 trials, it surpasses the accuracy upper bound of the NAS-Bench-201 search space, achieving 74.51% on CIFAR-100 (compared to the best in-space search result of 73.51%).
RL-RC-DoT: A Block-level RL Agent for Task-Aware Video Compression: Proposes RL-RC-DoT, a reinforcement learning-based macroblock-level quantization parameter (QP) control agent for task-aware video compression. By modeling QP selection as a sequential decision-making problem in RL, the agent learns to allocate more bitrate to task-relevant regions under given bitrate constraints, significantly improving performance on vehicle detection and ROI saliency coding tasks. A key advantage is that it does not require running downstream task models during inference, making it suitable for edge device deployment.
SceneAssistant: A Visual Feedback Agent for Open-Vocabulary 3D Scene Generation: Proposes SceneAssistant, a closed-loop agentic framework based on visual feedback. By designing a fully functioning suite of Action APIs (13 atomic operations spanning object search, deletion, 6DoF spatial operations, and camera control) for VLMs, this approach enables iterative, open-vocabulary 3D scene generation using the ReAct paradigm. It significantly outperforms Holodeck and SceneWeaver in both indoor (preference rate of 61.25%) and open-domain (preference rate of 65.00%) scenarios.
SpiritSight Agent: Advanced GUI Agent with One Look: This paper proposes SpiritSight, a vision-based end-to-end GUI agent, which resolves grounding ambiguity under dynamic high-resolution inputs through a multi-tier dataset of 5.73 million samples named GUI-Lasagne and the Universal Block Parsing (UBP) method. On Multimodal-Mind2Web under the non-candidate element setup, SpiritSight-8B achieves a Step Success Rate (SR) of 52.7%, outperforming all vision, language, and hybrid methods.
TANGO: Training-free Embodied AI Agents for Open-world Tasks: This paper proposes TANGO, which orchestrates two minimal navigation foundation primitives (PointGoal Navigation + Memory-based Exploration) through the program generation capability of LLMs. Without any task-specific training and using only few-shot examples, TANGO achieves state-of-the-art (SOTA) results across three distinct embodied AI tasks: Open-Set ObjectGoal Navigation, Multi-Modal Lifelong Navigation, and Open Embodied QA, demonstrating the generalizability of the "minimal primitive set + LLM composition" paradigm.
Visual Agentic AI for Spatial Reasoning with a Dynamic API: This paper proposes VADAR, an agentic program synthesis approach for 3D spatial reasoning. Multiple LLM agents collaborate to generate Pythonic APIs and dynamically extend new functions to solve common subproblems during the solving process, overcoming the limitations of prior methods like VisProg/ViperGPT that rely on static, human-defined APIs. At the same time, it introduces a new benchmark involving multi-step spatial localization and reasoning, outperforming existing zero-shot methods on 3D understanding tasks.

Browse all 12 LLM Agent papers →

⚖️ Alignment & RLHF (5)¶

CAD-Llama: Leveraging Large Language Models for Computer-Aided Design Parametric 3D Model Generation: This paper proposes the CAD-Llama framework, which converts 3D CAD models into Python-style code rich in semantic descriptions (SPCC) via a hierarchical annotation pipeline. It then utilizes adaptive pre-training and instruction tuning to transform LLaMA3-8B into a parametric CAD model generator. This approach outperforms previous methods by approximately 14% in accuracy on the text-to-CAD task, while supporting various CAD editing tasks such as completion, addition, and deletion.
Bases of Steerable Kernels for Equivariant CNNs: From 2D Rotations to the Lorentz Group: Proposes an alternative method to solve the constraint equations of steerable equivariant CNN kernels. By solving simpler invariance conditions at a fixed point and then "steering" to arbitrary points, this approach bypasses the need for computing Clebsch-Gordan coefficients, providing explicit kernel basis formulas for SO(2), O(2), SO(3), O(3), and the Lorentz group.
Continual SFT Matches Multimodal RLHF with Negative Supervision: Through gradient analysis, it is discovered that the core advantage of multimodal RLHF over continual SFT lies in the negative supervision signal within the rejected responses. Based on this, the nSFT method is proposed, which uses an LLM to extract error information from rejected responses and construct corrective dialogue data. It matches or even outperforms RLHF methods like DPO/PPO using only SFT loss, requiring only one model and significantly improving GPU memory efficiency.
Do We Really Need Curated Malicious Data for Safety Alignment in Multi-Modal LLMs?: This work investigates whether safety alignment in multimodal large language models genuinely requires meticulously curated malicious data. It demonstrates that effective safety alignment can be achieved by leveraging existing benign data combined with simple safety fine-tuning strategies, thereby significantly reducing the data curation cost associated with safety alignment.
Jailbreaking the Non-Transferable Barrier via Test-Time Data Disguising: This paper proposes JailNTL, the first black-box attack method against Non-Transferable Learning (NTL) models. By utilizing test-time data disguising to transform unauthorized domain data into the style of the authorized domain, it improves unauthorized domain accuracy by up to 55.7% using only 1% authorized samples, without requiring any model modifications.

🔒 Safety (46)¶

Hyperbolic Safety-Aware Vision-Language Models: HySAC proposes constructing safety-aware vision-language models in hyperbolic space. By mapping safe and unsafe content to different regions of hyperbolic space via entailment cones (safe content near the origin, unsafe content far from the origin), the model is equipped with safe content classification and dynamic redirection capabilities, significantly outperforming existing unlearning methods in retrieval safety and NSFW detection.
MP-GUI: Modality Perception with MLLMs for GUI Understanding: MP-GUI designs three specialized perceivers to extract graphical, textual, and spatial modal information from GUIs. By combining these three modalities through a spatial structure refinement strategy and an adaptive fusion gate, it outperforms general MLLMs on various GUI understanding tasks under limited training data.
Protecting Your Video Content: Disrupting Automated Video-Based LLM Annotations: This paper proposes two types of adversarial video watermarking methods—Ramblings (which induce video LLMs to generate incorrect descriptions) and Mutes (which induce video LLMs to generate extremely short or empty descriptions)—to protect personal videos from unauthorized automated annotation via imperceptible adversarial perturbations. It also demonstrates that these low-quality annotations degrade the performance of downstream text-to-video generation models.
Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks: This paper proposes ASTRA, which locates jailbreak-related visual tokens in adversarial images via image attribution, constructs steering vectors representing harmful response directions, and performs adaptive activation steering during inference to steer the model away from harmful directions. It achieves SOTA defense performance, with a 12% lower toxicity score, 18% lower ASR, and 9x faster speed compared to JailGuard.
A Closed-Form Solution for Debiasing Vision-Language Models with Utility Guarantees Across Modalities and Tasks: Proposed a training-free, data-free debiasing method for VLMs. By deriving closed-form solutions in a cross-modal space, it achieves Pareto-optimal trade-offs between fairness and utility retention, consistently outperforming existing approaches across three downstream tasks: zero-shot classification, text-to-image retrieval, and text-to-image generation.
CleanSight: Test-Time Attention Purification for Backdoored Large Vision Language Models: CleanSight reveals that the backdoor attack mechanism in LVLMs lies not in the pixel space but in the attention map—the trigger activates the backdoor via "attention stealing" (where trigger tokens hijack the attention of text tokens). Based on this, a training-free, plug-and-play test-time defense method is proposed: identifying poisoned inputs by detecting anomalies in cross-modal attention ratios, and neutralizing the backdoor by pruning high-attention visual tokens. This reduces the ASR to near 0% with almost no impact on model performance.
Dual Consolidation for Pre-Trained Model-Based Domain-Incremental Learning: This paper proposes the Duct method, which addresses exemplar-free domain-incremental learning on pre-trained models. Duct employs representation consolidation (accumulating task vectors to build a unified embedding space) and classifier consolidation (utilizing category semantic information via optimal transport to estimate the weights of classifiers for old domains). It outperforms state-of-the-art methods by 1% to 7% across four benchmark datasets.
ForensicZip: More Tokens are Better but Not Necessary in Forensic Vision-Language Models: Having discovered that semantic-driven visual token pruning discards forensic evidence (as tampering traces reside in low-saliency regions), this work proposes ForensicZip. It utilizes Birth-Death optimal transport to quantify physical inter-frame discontinuities and incorporates a high-frequency prior to preserve forensic signals. ForensicZip achieves 2.97x acceleration and 90%+ FLOPs reduction at a 10% token retention rate with no performance degradation.
LLM4SVG: Empowering LLMs to Understand and Generate Complex Vector Graphics: This paper proposes the LLM4SVG framework, which enables open-source LLMs (such as GPT-2, Phi-2, and Falcon) to understand and generate high-quality, complex vector graphics. This is achieved by defining 55 learnable SVG semantic tokens to replace raw XML tags and conducting a two-stage instruction fine-tuning process on the SVGX-SFT dataset, which contains 250K high-quality SVGs and 580K instruction-following pairs. The GPT-2 XL-based model achieves an FID of 64.11 and a CLIPScore of 0.3496, significantly outperforming GPT-4o (127.78 FID) and all existing SVG generation methods.
LoTUS: Large-Scale Machine Unlearning with a Taste of Uncertainty: This paper proposes LoTUS, which utilizes logit temperature adjustment and Gumbel-Softmax to smooth predictions of forgotten samples. By dynamically scheduling the temperature, it converges to the target where "forget set accuracy equals unseen set accuracy." This enables efficient unlearning on the large-scale ImageNet-1K benchmark (Avg Gap of 0.0150 on ViT). Furthermore, it introduces RF-JSD, a retraining-free evaluation metric (achieving a Pearson correlation of 0.92 with the true JSD).

Browse all 46 Safety papers →

👻 Hallucination Detection (9)¶

Seeing Far and Clearly: Mitigating Hallucinations in MLLMs with Attention Causal Decoding: FarSight is proposed as a plug-and-play, training-free decoding strategy. It introduces attention registers into the upper triangle matrix of the causal mask to absorb excessive attention on anomalous tokens, and designs positional awareness encoding with diminishing masking rates to enhance information propagation for distant visual tokens, thereby effectively mitigating initial and snowball hallucinations in Multimodal Large Language Models (MLLMs).
PhD: A ChatGPT-Prompted Visual Hallucination Evaluation Dataset: This paper proposes PhD, a large-scale visual hallucination evaluation dataset constructed with the assistance of ChatGPT. It contains 14K+ everyday images, 750 counter-commonsense images, and 102K VQA triplets. Through 4 evaluation modes $\times$ 5 visual tasks, it systematically evaluates the hallucination issues of multimodal large language models (MLLMs), far exceeding existing benchmarks in scale and difficulty.
3D-GRAND: A Million-Scale Dataset for 3D-LLMs with Better Grounding and Less Hallucination: This work constructs 3D-GRAND, the first million-scale densely grounded 3D scene-language dataset (40K scenes, 6.2M instructions), and proposes the 3D-POPE hallucination evaluation benchmark. It demonstrates that densely grounded instruction tuning significantly enhances the grounding capability of 3D-LLMs and reduces hallucinations, while also showcasing effective sim-to-real transfer.
Antidote: A Unified Framework for Mitigating LVLM Hallucinations in Counterfactual Presupposition and Object Perception: This paper proposes Antidote—a unified, synthetic data-driven post-training framework that enables model self-correction by injecting factual priors into prompts, decoupling hallucination mitigation as a preference optimization problem. It improves CP-Bench by over 50% on the LLaVA series, increases POPE by 1.8-3.3%, and reduces CHAIR/SHR by 30-50% without suffering from catastrophic forgetting.
HalLoc: Token-Level Localization of Hallucinations for Vision Language Models: This work proposes HalLoc, a token-level hallucination annotation dataset with 155K samples covering three categories of tasks: VQA, instruction following, and image captioning. Based on this, a lightweight hallucination detection model named HalLocalizer is trained, which can be integrated into existing VLMs in a plug-and-play manner to achieve real-time probabilistic hallucination detection without sacrificing efficiency.
Octopus: Alleviating Hallucination via Dynamic Contrastive Decoding: Through extensive experiments, this paper reveals the hybrid nature of hallucination causes in LVLMs—different samples and different generation steps face different flags of hallucination challenges. Consequently, the Octopus framework is proposed, which utilizes a learnable decision token and a transformer block to adaptively select the most appropriate contrastive decoding (CD) strategy at each generation step. Optimized via DPO, Octopus outperforms existing CD methods across four benchmarks.
ODE: Open-Set Evaluation of Hallucinations in Multimodal Large Language Models: This paper proposes the ODE (Open-set Dynamic Evaluation) protocol, which models real-world object concepts and their distribution associations using a graph structure. It dynamically extracts concept combinations and generates synthetic test images to realize open-set, continuously updated multimodal hallucination evaluation, effectively avoiding the data contamination issues potentially present in current static benchmarks.
One Token, Two Fates: A Unified Framework via Vision Token Manipulation Against MLLMs Hallucination: Proposes the first unified, training-free framework for mitigating MLLM hallucinations, operating synergistically within the hidden representation layers based on the dual roles of vision tokens—enhancement (SVC) and suppression (CRC). It improves POPE accuracy by ~2% on LLaVA-1.5 with only a 1.06× increase in inference latency.
Stop Learning It All to Mitigate Visual Hallucination, Focus on the Hallucination Target: Proposes TL-DPO (Target-Learning DPO), which limits traditional full-sentence preference learning to the target chunk where hallucination occurs and the corresponding image region. By excluding irrelevant signals through target generation loss and target condition loss, it reduces CHAIR_s on LLaVA-1.5 from 66.8 to 20.1, while improving LLaVA-Bench from 63.4 to 71.2.

📊 LLM Evaluation (4)¶

PosterO: Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation: PosterO is proposed to structure poster layouts into SVG layout trees. By vectorizing design intents and modeling hierarchical node representations, it interfaces with LLMs, generating high-quality content-aware layouts via intent-aligned in-context learning. It achieves state-of-the-art performance across multiple benchmarks and introduces the first PStylish7 dataset supporting multi-purpose and multi-shape elements.
Erase Diffusion: Empowering Object Removal Through Calibrating Diffusion Pathways (EraDiff): This paper proposes EraDiff, which establishes a progressive diffusion pathway from "object-containing" to "pure background" through the Chain-Rectifying Optimization (CRO) paradigm, and suppresses artifacts during sampling using the Self-Rectifying Attention (SRA) mechanism. This enables the diffusion model to truly comprehend the "erasure intention," achieving a SOTA Local FID (3.799) on OpenImages V5 and significantly outperforming SD2-Inpaint and LaMa in complex real-world scenes.
UniGoal: Towards Universal Zero-shot Goal-oriented Navigation: This paper proposes UniGoal, a unified zero-shot goal-oriented navigation framework. By representing both scenes and goals uniformly as graph structures and combined with a graph matching-driven multi-stage exploration strategy, it achieves zero-shot navigation for three goal types—object categories, instance images, and text descriptions—within a single model, outperforming task-specific methods.
RoadSocial: A Diverse VideoQA Dataset and Benchmark for Road Event Understanding from Social Video Narratives: This paper proposes RoadSocial, a large-scale and diverse VideoQA dataset sourced from social media (consisting of 13.2K videos and 260K QA pairs) that covers multi-regional and multi-perspective road event scenarios globally. Through a semi-automatic annotation framework and 12 categories of QA tasks, the paper systematically evaluates the road event understanding capabilities of 18 Video LLMs.

⚡ LLM Efficiency (5)¶

Spatial-TTT: Streaming Visual-based Spatial Intelligence with Test-Time Training: This paper proposes Spatial-TTT, which leverages the Test-Time Training (TTT) mechanism to utilize a subset of model parameters (fast weights) as compact non-linear memory. Combined with a hybrid architecture and a spatial prediction mechanism, the model continuously accumulates and organizes 3D spatial evidence from unbounded video streams, achieving SOTA on video spatial understanding benchmarks.
Associative Transformer: The Associative Transformer (AiT) is proposed, which integrates a learnable explicit memory module and a Hopfield network for token reconstruction within the Transformer architecture, achieving classification and relational reasoning performance superior to ViT with fewer parameters.
Efficient Data Driven Mixture-of-Expert Extraction from Trained Networks: A post-training method is proposed to extract MoE variants from pre-trained ViTs. By automatically discovering expert structures using HDBSCAN to cluster MLP hidden layer activation patterns, it reduces MACs by 36% and parameters by 32% on ImageNet-1k while preserving 98% of the original accuracy without retraining.
LOCORE: Image Re-ranking with Long-Context Sequence Modeling: This paper proposes LoCoRe (Long-Context Re-ranker), achieving list-wise image re-ranking based on local descriptors for the first time. By leveraging the Longformer long-context sequence model to process the local descriptors of both the query image and the entire candidate list simultaneously, LoCoRe significantly improves re-ranking performance by capturing transitive relations among candidate images.
Efficient Data Driven Mixture-of-Expert Extraction from Trained Networks: A method is proposed to automatically extract MoE (Mixture-of-Experts) variants from pre-trained ViTs. By first clustering the output activation patterns of MLP layers and then extracting corresponding subnetworks as experts, this approach avoids training MoEs from scratch. It recovers 98% of the original performance on ImageNet-1k with only minimal fine-tuning, while reducing FLOPs and model size by 36% and 32%, respectively.

📚 Pretraining (15)¶

MR-PLIP: Multi-Resolution Pathology-Language Pre-training Model with Text-Guided Visual Representation: Proposed MR-PLIP, the first multi-resolution pathology vision-language pre-training model. Pre-trained on 34 million multi-resolution image-text pairs from the TCGA dataset, it outperforms SOTA on 26 datasets through cross-resolution vision-text alignment and text-guided visual representation.
ConText-CIR: Learning from Concepts in Text for Composed Image Retrieval: The ConText-CIR framework is proposed, which utilizes a Text Concept-Consistency loss to align noun phrases in text modifications with corresponding regions in the query image. Combined with a synthetic data generation pipeline, it achieves SOTA performance on multiple CIR benchmarks.
A Unified Framework for Heterogeneous Semi-supervised Learning: This paper proposes a new problem setting termed Heterogeneous Semi-Supervised Learning (HSSL), where labeled and unlabeled data originate from domains with different distributions, and the goal is to train a model that generalizes well to both domains. By expanding the C-class problem into a 2C-class classification task (where the same semantic class in different domains is treated as distinct classes), this work provides a unified solution integrating Weighted Moving Average (WMA) pseudo-labeling, cross-domain prototype alignment, and progressive cross-domain Mixup.
AMO Sampler: Enhancing Text Rendering with Overshooting: This paper proposes the Attention-Modulated Overshooting (AMO) sampler, a training-free inference-time enhancement method. By introducing an overshooting-noise compensation Langevin dynamics correction during the sampling process of rectified flow models, and adaptively controlling the overshooting intensity using text-image cross-attention scores, it significantly improves text rendering accuracy while maintaining the overall quality of generated images.
Bridging the Vision-Brain Gap with an Uncertainty-Aware Blur Prior: This work introduces the concepts of "System GAP" and "Random GAP" for the first time to describe the information mismatch between brain signals and visual stimuli. By dynamically adjusting the image blur level through an Uncertainty-Aware Blur Prior (UBP) to alleviate overfitting during training, it achieves a 50.9% top-1 accuracy on the 200-way zero-shot brain-image retrieval task, outperforming the previous SOTA by 13.7 percentage points.
DreamText: High Fidelity Scene Text Synthesis: DreamText reconstructs the training pipeline of diffusion models, introducing character-level balanced supervision and a heuristic alternate optimization strategy to calibrate character attention. Combined with the joint training of the text encoder and generator to learn diverse font styles, it significantly outperforms state-of-the-art methods in scene text synthesis tasks (improving SeqAcc from 0.763 of UDiffText to 0.940).
Exploration-Driven Generative Interactive Environments: This work provides an open-source implementation of the Genie world model (GenieRedux), which is enhanced to GenieRedux-G by incorporating ground-truth action conditioning, Token Distance Cross-Entropy (TDCE) loss, and token skip connections. Additionally, the AutoExplore agent is proposed to utilize the world model's token prediction uncertainty as an intrinsic reward to drive diverse data collection, improving simulation quality by up to 7.4 PSNR.
Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction: This paper proposes IAR, which rearranges the VQGAN codebook via balanced K-means to align similar embeddings with adjacent indices. Combined with a cluster-oriented cross-entropy loss that guides the model to correctly predict the semantic cluster of the target token, IAR halves the training time while improving generation quality across all LlamaGen scales from 100M to 1.4B.
Influence Malleability in Linearized Attention: Dual Implications of Non-Convergent NTK Dynamics: Reveals through the NTK framework that linearized attention mechanisms do not converge to the infinite-width NTK limit (the spectral amplification effect cubes the condition number of the Gram matrix, requiring a width of $m = \Omega(\kappa^6)$), and introduces the concept of "influence malleability" to quantify the dual consequences of this non-convergence: an attention network's malleability, which is 6-9 times higher than that of a ReLU network, both enhances task adaptability and exacerbates adversarial vulnerability.
PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes: This paper proposes PlanarSplatting, which directly optimizes learnable 3D rectangular plane primitives. By utilizing a newly designed rectangular splatting function, planes are differentiably rendered into depth and normal maps. This enables the reconstruction of accurate indoor planar scenes from multi-view images in just 3 minutes without requiring any plane annotations.

Browse all 15 Pretraining papers →

✏️ Knowledge Editing (1)¶

MoKus: Leveraging Cross-Modal Knowledge Transfer for Knowledge-Aware Concept Customization: Proposes the MoKus framework, discovering and utilizing the "cross-modal knowledge transfer" phenomenon—where updating knowledge in an LLM text encoder propagates automatically to the visual generation end—to achieve knowledge-aware concept customization. It features a two-stage design: first learning a visual anchor representation, and then binding textual knowledge in seconds.

💬 LLM (Other) (15)¶

MG-MotionLLM: A Unified Framework for Motion Comprehension and Generation across Multiple Granularities: MG-MotionLLM proposes a unified multi-granularity motion-language model. Leveraging a Motion VQ-VAE + T5 language model architecture along with a carefully designed multi-granularity synergy pre-training scheme (comprising 28 tasks), it simultaneously supports coarse- and fine-grained motion comprehension and generation. While achieving state-of-the-art performance on classic tasks, it also enables novel applications such as fine-grained motion editing.
Building Vision Models upon Heat Conduction: A vision backbone named vHeat is proposed, which models image patches as heat sources and utilizes physical heat conduction equations via DCT/IDCT transforms to achieve global information propagation with $O(N^{1.5})$ complexity. It achieves 84.0% top-1 accuracy on ImageNet-1K with 3x higher throughput and 80% less GPU memory overhead.
Imagine and Seek: Improving Composed Image Retrieval with an Imagined Proxy: Proposed the IP-CIR method, which translates Composed Image Retrieval (CIR) into a standard image retrieval problem by using large language models to generate an "imagined target text description" as a proxy, achieving zero-shot SOTA on benchmarks such as CIRR and FashionIQ.
ComRoPE: Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices: This paper proposes ComRoPE, which generalizes RoPE into a rotary position embedding parameterized by trainable commuting angle matrices. It theoretically proves that the pairwise commutativity of angle matrices is a necessary and sufficient condition for RoPE to satisfy relative position dependency, outperforming the state-of-the-art LieRE method by 1.6% (on training resolution) and 2.9% (on higher resolutions) on ImageNet-1K.
Dora: Sampling and Benchmarking for 3D Shape Variational Auto-Encoders: Proposes Dora-VAE, which focuses on sharp geometric edge regions via Sharp Edge Sampling (SES) and handles uniform and salient sampled points separately using Dual Cross-Attention. It achieves superior 3D shape reconstruction quality with only 1,280 latent codes (8× fewer than XCube-VAE's 10,000+), while establishing a new evaluation benchmark, Dora-Bench.
Spiking Transformer with Spatial-Temporal Attention: Spatially-temporally decoupled attention is integrated into the Spiking Transformer architecture. By combining spatial-temporal decoupled attention designs with a spike-driven self-attention mechanism, the approach bridges the performance gap with ANNs while preserving the energy efficiency advantages of SNNs, achieving SOTA performance on multiple vision benchmarks.
STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks: By integrating four major modules—Global Context self-attention (GC), Position Encoding (PE), Step Attention (SA), and Timestep Random Dropout (TSRD)—into SNNs, STAA-SNN achieves state-of-the-art (SOTA) performance on CIFAR-10/100 and ImageNet with accuracies of 97.14%/82.05%/70.40%, respectively.
Test-Time Visual In-Context Tuning: This paper proposes Visual In-Context Tuning (VICT), which performs one-shot adaptation of visual in-context learning models (e.g., Painter) at test time by flipping the roles of task prompts and test samples and utilizing cycle consistency loss, significantly improving generalization under distribution shifts.
Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment: This paper proposes a new paradigm of Chat-based Person Retrieval (ChatPR), builds the first dialogue-image paired dataset ChatPedes, and designs the DiaNA framework to achieve fine-grained cross-modal alignment between dialogues and images via an adaptive attribute refiner, significantly outperforming traditional single-sentence text retrieval methods.
Exposure-slot: Exposure-centric Representations Learning with Slot-in-Slot Attention: This paper proposes the Exposure-slot framework, which extends the Slot Attention algorithm into a hierarchical slot-in-slot structure. Guided by learnable exposure prompts for feature clustering, it achieves exposure-centric region-aware representation learning, obtaining SOTA performance in under-/over-exposed image correction tasks.

Browse all 15 LLM (Other) papers →

✍️ Text Generation (2)¶

ArtFormer: Controllable Generation of Diverse 3D Articulated Objects: This work proposes the ArtFormer framework, which generates high-quality, diverse, and kinematically accurate 3D articulated objects from text/image descriptions via tree structure parameterization and a conditional diffusion shape prior, significantly outperforming existing methods in generation quality and diversity.
Dense Match Summarization for Faster Two-view Estimation: This paper proposes a dense match summarization scheme that compresses over 10,000 dense matches into approximately 1% representative matches through clustering and representative match selection. It encodes the geometric constraints of each cluster into a 9×9 matrix, achieving a 10× to 100× speedup for robust RANSAC estimation with negligible accuracy loss.

🌐 Multilingual & Translation (1)¶

SMTPD: A New Benchmark for Temporal Prediction of Social Media Popularity: Constructs SMTPD, the first time-aligned temporal prediction benchmark for social media popularity (282K YouTube samples with 30-day continuous observations), proposes a baseline framework based on multi-modal feature extraction and LSTM temporal regression, and reveals that early popularity (EP) is key to accurately predicting subsequent popularity.

🔍 Information Retrieval & RAG (12)¶

Retrieving Semantics from the Deep: an RAG Solution for Gesture Synthesis: RAG-Gesture proposes a gesture synthesis framework based on Retrieval-Augmented Generation (RAG). It leverages explicit linguistic knowledge to retrieve semantically relevant exemplar motions from a gesture database, and injects them into the diffusion model's generation process at inference time through DDIM inversion and retrieval guidance, producing semantically rich and natural co-speech gestures without training.
GOAL: Global-Local Object Alignment Learning: Proposes the GOAL method, which enhances CLIP's understanding of long text descriptions through two modules: Local Image-Sentence Matching (LISM) and Token Similarity-based Learning (TSL). By introducing local semantic alignment on top of global alignment, it significantly improves image-text retrieval performance.
ChatHuman: Chatting about 3D Humans with Tools: ChatHuman is proposed, an LLM-based language-driven system that manages new tools by automatically selecting and integrating specialized 3D human analysis tools (3D pose estimation, shape recovery, contact detection, human-object interaction analysis, emotion recognition, etc.), utilizing academic papers as tool manuals along with RAG (Retrieval-Augmented Generation) to create in-context examples. It outperforms existing LLM models in tool selection accuracy and overall performance on 3D human-related tasks.
VDocRAG: Retrieval-Augmented Generation over Visually-Rich Documents: Proposes the first RAG framework that directly takes document images (rather than parsed text) as input, utilizing an LVLM as a dual-encoder retriever plus two self-supervised pre-training tasks (contrastive + generative) to achieve document image retrieval, outperforming text RAG by 24 points on ChartQA.
Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning: This paper extends Retrieval-Augmented Learning (RAL) to Few-Shot Recognition (FSR) for the first time, exposing two main challenges of retrieved data: distribution imbalance and domain gap. It proposes a two-stage method, SWAT (finetuning the vision encoder on mixed data first, then retraining the classifier on few-shot labeled data), outperforming all prior methods by $>6\%$ across 9 benchmarks.
Advancing Myopia To Holism: Fully Contrastive Language-Image Pre-training: Upgrades CLIP from the traditional one-to-one (image, text) contrastive learning paradigm to a multi-to-multi (multi-image-embeddings, multi-texts) contrastive learning paradigm. By utilizing VLMs to generate multi-perspective, multi-level descriptions and a multi-branch visual encoder to output diverse visual embeddings, it achieves more comprehensive vision-language alignment, substantially outperforming baselines in retrieval, classification, and dense prediction tasks.
COBRA: COmBinatorial Retrieval Augmentation for Few-Shot Adaptation: Proposes COBRA, a combinatorial mutual information (CMI)-based retrieval-augmented few-shot adaptation method. By simultaneously accounting for both the similarity of retrieved samples to the target task and the diversity among the samples themselves, COBRA retrieves high-quality auxiliary data from LAION-2B. It consistently outperforms traditional nearest-neighbor retrieval methods across multiple image classification benchmarks with negligible computational overhead.
EZSR: Event-based Zero-Shot Recognition: This paper proposes the EZSR framework for zero-shot object recognition in event camera data. By utilizing a scalar-wise modulation strategy, it addresses the semantic misalignment between event embeddings and CLIP text embeddings. It overcomes training data scarcity through large-scale event data synthesis from static RGB images, achieving a 47.84% zero-shot accuracy on N-ImageNet with a ViT-B/16 backbone.
LotusFilter: Fast Diverse Nearest Neighbor Search via a Learned Cutoff Table: This paper proposes LotusFilter, which constructs a cutoff table by precomputes neighbor relationships for each vector offline and performs diversity filtering using greedy set deletion during the online stage. This reduces the complexity of traditional diverse search from $O(DS^2)$ to $O(T+S+KL)$. The filtering process requires only 0.02 ms/query, utilizing only 1/40 of the memory compared to traditional methods.
Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation: The CRPL framework is proposed to improve prompt learning of CLIP in unsupervised domain adaptation (UDA) through source-augmented pseudo-labeling and an optimal transport-based cluster preservation strategy, ensuring that the text embeddings of target prompts better cover the cluster structures of visual embeddings.

Browse all 12 Information Retrieval & RAG papers →

🎨 Image Generation (313)¶

Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing: This paper proposes Decoupled Annealed Posterior Sampling (DAPS). By decoupling the sample dependencies between adjacent steps during the diffusion sampling process, it allows large-scale non-local jumps to correct early sampling errors, substantially outperforming existing methods on non-linear inverse problems (e.g., phase retrieval).
Stable Flow: Vital Layers for Training-Free Image Editing: Stable Flow proposes to automatically detect "vital layers" in DiT (FLUX) and inject attention features of the reference image only into these layers to achieve various training-free image editing operations, while introducing a latent nudging technique to improve the quality of flow model inversion for real images.
Diffusion-4K: Ultra-High-Resolution Image Synthesis with Latent Diffusion Models: This paper proposes the Diffusion-4K framework, which consists of the Aesthetic-4K benchmark dataset, GLCM Score/Compression Ratio evaluation metrics, and a wavelet-based fine-tuning method. It enables large-scale latent diffusion models (LDMs) such as SD3-2B and Flux-12B to directly generate high-quality 4096×4096 images with rich texture details.
Co-Spy: Combining Semantic and Pixel Features to Detect Synthetic Images by AI: Co-Spy is proposed to fuse two complementary detection pathways: VAE reconstruction artifact features and CLIP semantic features. VAE artifacts generalize well across models but are vulnerable to JPEG compression, while CLIP semantics are robust to JPEG compression but generalize poorly. An adaptive regulator dynamically allocates weights to the two pathways based on the input, establishing a new SOTA across 22 generative models.
SALAD: Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing: This work proposes SALAD, a skeleton-aware latent diffusion model that explicitly models fine-grained interactions among joints, frames, and text using a skeleton-temporal structured VAE and denoiser, and achieves zero-shot text-driven motion editing via cross-attention maps.
OSDFace: One-Step Diffusion Model for Face Restoration: OSDFace proposes the first one-step diffusion model specifically designed for face restoration. By extracting rich prior information from low-quality faces using a Visual Representation Embedder (VRE), and combining this with facial identity loss and GAN guidance, it generates high-fidelity, natural, and identity-consistent face images in just a single step of inference (~0.1 second), comprehensively outperforming existing SOTA methods.
Dissecting and Mitigating Diffusion Bias via Mechanistic Interpretability: This paper proposes the DiffLens framework, which disentangles the internal neurons of diffusion models into a monosemantic feature space using sparse autoencoders (k-SAEs), and then localizes specific features driving bias generation through gradient-based attribution. This enables fine-grained control and mitigation of social biases such as gender and race while maintaining image quality.
FineLIP: Extending CLIP's Reach via Fine-Grained Alignment with Longer Text Inputs: This paper proposes FineLIP, which supports long text inputs up to 248 tokens via position embedding stretching and introduces adaptive token refinement alongside cross-modal token-level alignment, significantly outperforming state-of-the-art methods in long-description text retrieval and text-to-image generation tasks.
Generative Multimodal Pretraining with Discrete Diffusion Timestep Tokens: DDT-LLaMA proposes using diffusion timestep encoding to learn discrete visual tokens (DDT) with a recursive structure. This endows visual token sequences with hierarchical dependencies similar to natural language, thereby achieving SOTA performance in both multimodal understanding and generation under a unified next-token-prediction framework.
Personalized Preference Fine-tuning of Diffusion Models: PPD proposes a personalized preference diffusion model fine-tuning framework: it leverages a VLM to extract user embeddings from a small number of (4 pairs) preference examples, injects them into the diffusion model via decoupled cross-attention layers, and optimizes multi-user personalized preferences simultaneously using a DPO objective. It requires only 4 preference pairs to generate preference-aligned images for a new user (76% win rate).

Browse all 313 Image Generation papers →

🎬 Video Generation (89)¶

GEN3C: 3D-Informed World-Consistent Video Generation with Precise Camera Control: GEN3C proposes a video generation framework guided by a 3D cache (point cloud cache). By predicting depth for seed images and unprojecting them into 3D point clouds, it renders the 3D cache into 2D condition maps according to user-specified camera trajectories when generating subsequent frames, thereby achieving precise camera control and cross-frame 3D consistency.
HunyuanPortrait: Implicit Condition Control for Enhanced Portrait Animation: HunyuanPortrait proposes the first implicit condition portrait animation framework based on Stable Video Diffusion, achieving high-fidelity control of fine facial dynamics and robust identity consistency through an intensity-aware motion encoder and an ID-aware multi-scale adapter.
Video-ColBERT: Contextualized Late Interaction for Text-to-Video Retrieval: This paper proposes Video-ColBERT, which introduces ColBERT's late interaction in text retrieval into text-to-video retrieval (T2VR). By performing MeanMaxSim interaction at both the frame and video levels and employing a dual Sigmoid loss to train independent yet compatible multi-granularity representations, Video-ColBERT outperforms existing dual-encoder methods on multiple T2VR benchmarks.
Teller: Real-Time Streaming Audio-Driven Portrait Animation with Autoregressive Motion Generation: This work proposes Teller, the first real-time streaming audio-driven portrait animation framework based on an autoregressive Transformer. By leveraging RVQ to discretize facial motion into tokens and combining it with an efficient temporal module to refine body details, Teller achieves 25 FPS real-time generation speed (requiring only 0.92s to generate a 1s video, compared to 20.93s with Hallo) while delivering animation quality comparable to diffusion models.
LongDiff: Training-Free Long Video Generation in One Go: Through theoretical analysis, LongDiff reveals two key challenges when short-video models generate long videos: temporal position blurring and information dilution. It proposes two simple temporal attention modification strategies, Position Mapping (GROUP+SHIFT) and Informative Frame Selection (IFS), enabling short-video models to generate high-quality long videos in one go without training.
Articulated Kinematics Distillation from Video Diffusion Models: This paper proposes the AKD framework, which reduces the degrees of freedom of 3D asset motion from full space to a small number of joint angles through skeletal joint parameterization, then distills text-aligned joint motion sequences using SDS gradients from a video diffusion model (CogVideoX), and further ensures physical plausibility via physical simulation.
Learning from Streaming Video with Orthogonal Gradients: Addressing the issue of gradient redundancy and model collapse caused by highly correlated continuous frames in streaming video learning, an Orthogonal Optimizer is proposed. By projecting the current gradient onto the orthogonal component of historical gradients for decorrelation, it can be seamlessly integrated into SGD/AdamW. It significantly recovers the performance loss of transitioning from shuffled training to sequential training across three scenarios: DoRA, VideoMAE, and future prediction.
Visual Prompting for One-Shot Controllable Video Editing Without Inversion: This work tackles the One-Shot Controllable Video Editing (OCVE) problem from a novel perspective of Visual Prompting. By leveraging an image inpainting diffusion model to perform editing propagation, and introducing Content Consistency Sampling (CCS) and Temporal-Content Consistency Sampling (TCS), the method achieves high-quality controllable video editing without DDIM inversion.
Out of Sight, Out of Mind? Evaluating State Evolution in Video World Models: StEvo-Bench proposes a benchmark to evaluate the capabilities of video world models in "unobserved state evolution"—testing whether world models can continue to correctly reason about state changes when physical processes are unobserved (due to camera movement, occlusion, or turning off lights). The results reveal a severe "out of sight, out of mind" deficiency, with all current frontier models (e.g., Veo 3, Sora 2 Pro) achieving task success rates of less than 10%.
Zero-1-to-A: Zero-Shot One Image to Animatable Head Avatars Using Video Diffusion: Zero-1-to-A is proposed to generate high-fidelity animatable 4D head avatars from a single image using a pretrained video diffusion model via Symbiotic Generation (SymGEN) and a progressive learning strategy, effectively addressing the spatio-temporal inconsistency issue in video diffusion.

Browse all 89 Video Generation papers →

🧩 Multimodal VLM (137)¶

Your Large Vision-Language Model Only Needs a Few Attention Heads for Visual Grounding: It is discovered that frozen LVLMs naturally contain a small number of "localization heads" that consistently capture object locations corresponding to textual semantics. Using the attention maps of only 3 attention heads, training-free visual grounding achieves 86.5% on RefCOCO val, outperforming the fine-tuned LISA-7B.
Words or Vision: Do Vision-Language Models Have Blind Faith in Text?: This paper identifies the phenomenon of "blind faith in text" in VLMs—where models systematically favor text (even when incorrect) when visual and textual inputs are inconsistent. By constructing a benchmark with three text variants (Match, Corruption, and Irrelevance), this work evaluates 10 VLMs, analyzes five influencing factors, demonstrates that SFT with text augmentation effectively mitigates this issue, and provides a theoretical explanation tracing the root cause to the imbalance between text-only and multimodal training data.
Rethinking Vision-Language Model in Face Forensics: Multi-Modal Interpretable Forged Face Detector: Proposed M2F2-Det, the first multimodal face forgery detector that simultaneously outputs deepfake detection scores and textual explanations. It adapts CLIP to learn forgery features via Forgery Prompt Learning, fuses CLIP and deepfake encoder features using a Bridge Adapter, and guides the LLM to generate trustworthy explanations using frequency-domain tokens.
Global-Local Tree Search in VLMs for 3D Indoor Scene Generation: Proposes a global-local tree search algorithm that leverages the spatial reasoning capabilities of VLMs. Through hierarchical scene representations and visual prompts from an emoji grid, it achieves high-quality 3D indoor scene layout generation, ranking first on average in user studies.
DocoPilot: Improving Multimodal Models for Document-Level Understanding: This paper constructs Doc-750K—a high-quality, document-level multimodal dataset containing 758K question-answer pairs and 3.1M images. Based on this, the authors train Docopilot, a native document understanding model. It outperforms InternVL2-8B by 19.9 percentage points on MM-NIAH, processing multi-page documents efficiently without relying on RAG.
MARTEN: Visual Question Answering with Mask Generation for Multi-Modal Document Understanding: The VQAMask pre-training paradigm is proposed, which introduces an auxiliary mask generation task (discarded during inference) on top of VQA text parsing. Through explicit spatial alignment supervision, it enhances the vision encoder's perception of text regions in document images. The resulting Marten model achieves state-of-the-art (SOTA) performance among 8B-level MLLMs across multiple document understanding tasks.
MIMO: A Medical Vision Language Model with Visual Referring Multimodal Input and Pixel Grounding Multimodal Output: This paper proposes MIMO, the first medical vision-language model that simultaneously supports "visual referring multimodal input" (users specify regions of interest via points/boxes) and "pixel-level grounding multimodal output" (the model embeds segmentation masks within textual answers). It constructs the MIMOSeg dataset with 895K samples and demonstrates unique referring + grounding capabilities across various medical VQA and segmentation tasks.
Dynamic Updates for Language Adaptation in Visual-Language Tracking: DUTrack is proposed to resolve the semantic inconsistency between static references and dynamic targets in visual-language tracking by dynamically updating multi-modal reference information (template frames + language descriptions), outperforming the best vision-only trackers on LaSOT for the first time.
Identifying and Mitigating Position Bias of Multi-image Vision-Language Models: This paper identifies a severe position bias in multi-image Large Vision-Language Models (LVLMs)—where open-source models place excessive emphasis on trailing images and closed-source models neglect middle images. It proposes a training-free SoFt Attention (SoFA) method that mitigates this bias by linearly interpolating between causal attention and bidirectional attention across images, improving average accuracy by 2~3% across multiple benchmarks.
DPC: Dual-Prompt Collaboration for Tuning Vision-Language Models: This work proposes the Dual-Prompt Collaboration (DPC) framework. By freezing the original tuned prompt to maintain new-class generalization and training a parallel prompt to strengthen base-class performance, along with a weighted decoupled inference mechanism, DPC serves as a plug-and-play module that consistently improves the base-new harmonic mean across four prompt tuning baselines.

Browse all 137 Multimodal VLM papers →

🧠 VLM Reasoning (14)¶

MV-MATH: Evaluating Multimodal Math Reasoning in Multi-Visual Contexts: This paper proposes the MV-MATH benchmark, consisting of 2,009 high-quality multi-image math problems (sourced from real K-12 scenarios) to systematically evaluate the capability of 25 multimodal large models in multi-image math reasoning scenarios. It is found that all models perform well below human levels (the best, Claude, only achieves 33.9%), revealing that multi-image math reasoning remains a significant challenge for MLLMs.
Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces: This paper proposes VSI-Bench, a video-based visual spatial intelligence benchmark (5000+ QA pairs), to systematically evaluate the spatial reasoning capabilities of MLLMs. The study finds that spatial reasoning is the primary bottleneck, and traditional language reasoning techniques (such as CoT) fail to improve performance, whereas explicitly generating cognitive maps can enhance spatial distance reasoning.
Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Models: This paper proposes Coarse Correspondences, a lightweight, training-free visual prompting method. By overlaying coarse-grained instance correspondence markers obtained from object tracking onto image frames, it significantly enhances the spatial-temporal reasoning capabilities of MLLMs, achieving improvements of +20.5% on ScanQA, +9.7% on OpenEQA, +6.0% on EgoSchema, and +11% on R2R navigation.
Critic-V: VLM Critics Help Catch VLM Errors in Multimodal Reasoning: This paper proposes the Critic-V framework, which decouples the VLM reasoning process into a Reasoner and a Critic. By utilizing a DPO-trained Critic model to provide natural language feedback for iteratively optimizing the reasoning path, this approach outperforms GPT-4V on 5 out of 8 benchmarks, showing particularly significant improvements on mathematical reasoning tasks (MathVista +11.8%).
Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models: Insight-V proposes a visual reasoning enhancement scheme consisting of a data generation pipeline and a multi-agent reasoning system: it constructs high-quality long-chain reasoning data through progressive generation and multi-granular evaluation, designs a Reasoning Agent and a Summary Agent to collaboratively solve problems, and incorporates iterative DPO to further improve reasoning quality, achieving an average improvement of 7% across seven visual reasoning benchmarks.
Beyond Final Answers: CRYSTAL Benchmark for Transparent Multimodal Reasoning Evaluation: This paper proposes the CRYSTAL benchmark (6372 instances) to evaluate MLLMs at the intermediate reasoning step level using Match F1 and Ordered Match F1. It reveals widespread cherry-picking behaviors and disordered reasoning processes, and introduces a CPR-Curriculum training strategy to improve reasoning quality.
ESPIRE: A Diagnostic Benchmark for Embodied Spatial Reasoning of Vision-Language Models: This paper proposes Espire, a simulation-based diagnostic benchmark for embodied spatial reasoning. It decomposes VLM evaluation into localization and execution phases, systematically assessing the capabilities of VLMs across multiple spatial reasoning dimensions and granularities through a fully generative paradigm.
MM-CondChain: A Programmatically Verified Benchmark for Visually Grounded Deep Compositional Reasoning: MM-CondChain is the first MLLM benchmark for visually grounded deep compositional reasoning. By using a Verifiable Programmatic Intermediate Representation (VPIR), it automatically constructs multi-layer conditional chains and chain-style hard negatives. The strongest model achieves only a 53.33 Path F1, revealing that deep compositional reasoning remains a fundamental challenge.
Reasoning over Video: Evaluating How MLLMs Extract, Integrate, and Reconstruct Spatiotemporal Evidence: Proposes the VAEX-Bench benchmark to systematically evaluate the "abstract spatiotemporal reasoning" capability of MLLMs for the first time. Unlike extractive tasks that pull information from single frames, abstract reasoning requires integrating observations across rooms and time to infer global spatial layouts, perform cross-scene counting, etc. The study reveals that all SOTA models (including GPT-5.2 and Gemini-3 Pro) perform significantly worse than humans on abstract reasoning.
SeqAfford: Sequential 3D Affordance Reasoning via Multimodal Large Language Model: This work proposes the Sequential 3D Affordance Reasoning task and constructs a benchmark of 180K instruction-point cloud pairs. By introducing a <SEG> token and a multi-granular language-point integration module into a 3D MLLM, the model reasons and segments sequential affordance regions from complex human instructions.

Browse all 14 VLM Reasoning papers →

⚡ VLM Efficiency (3)¶

MBQ: Modality-Balanced Quantization for Large Vision-Language Models: This work identifies that the sensitivity of vision tokens and language tokens to quantization errors in large VLMs differs by more than tenfold. It proposes MBQ, a post-training quantization method that introduces a gradient-based modality-balancing factor during calibration. Under W3A16 and W4A8 configurations, MBQ improves accuracy by up to 4.4% and 11.6%, respectively, while achieving a 1.4× end-to-end acceleration.
Quantization without Tears: This paper proposes the QwT (Quantization without Tears) method, which compensates for quantization information loss by adding a lightweight linear compensation layer after each block of the quantized network. The parameters of this compensation layer can be obtained via a closed-form solution in under 2 minutes, significantly improving PTQ accuracy across various tasks including vision, language, and multimodality.

COAP: Memory-Efficient Training with Correlation-Aware Gradient Projection

🎵 Audio & Speech (21)¶

Crab: A Unified Audio-Visual Scene Understanding Model with Explicit Cooperation: This paper proposes Crab, a unified audio-visual scene understanding model. By constructing the AV-UIE dataset (200K samples) with explicit reasoning processes, it clarifies the collaborative relationships across tasks. Combined with interaction-aware LoRA (multi-head LoRA) designed to learn different audio-visual interaction patterns, Crab outperforms specialized models across multiple tasks.
Object-aware Sound Source Localization via Audio-Visual Scene Understanding: This paper proposes OA-SSL: during the training phase, an MLLM is used to generate fine-grained descriptions of "$K$ sounding objects + 1 silent object" for each image as additional supervision anchors. Then, OCA (object-aware contrastive alignment) and ORI (object region isolation) losses are employed, enabling the model to locate only the truly sounding objects even in complex scenarios with multiple guitars where only one is being played.
Synchronized Video-to-Audio Generation via Mel Quantization-Continuum Decomposition: Mel-QCD is proposed to decompose Mel spectrograms into three signals: semantic vectors (quantized), energy, and standard deviation (continuous). By predicting these signals from video via a V2X predictor, and combining ControlNet with textual inversion, this approach achieves comprehensive SOTA video-to-audio generation across eight metrics on VGGSound.
ImViD: Immersive Volumetric Videos for Enhanced VR Engagement: This work constructs the first immersive volumetric video dataset by capturing 7 indoor/outdoor scenes using a mobile multi-view system with 46 synchronized GoPros. It proposes STG++, which introduces learnable affine color transformations to resolve cross-camera color inconsistency, achieving rendering at 110.47 FPS with 387MB of storage, and integrates HRTF spatial audio.
Towards Lossless Implicit Neural Representation via Bit Plane Decomposition: Discovers that the model capacity upper bound of implicit neural representation (INR) grows exponentially with bit precision ($\mathcal{P}(f_\theta) \propto 2^n$), and proposes bit-plane decomposition—decomposing an n-bit signal into n independent 1-bit planes to train individual INRs, achieving lossless (BER=0) implicit neural representation of 16-bit images for the first time.
Improving Sound Source Localization with Joint Slot Attention on Image and Audio: Proposes a joint slot attention mechanism to decompose both images and audio into target/non-target representations, achieving precise sound source localization through cross-modal attention matching and contrastive learning, resulting in SOTA performance of 65.16% AUC and 86.00% cIoU on Flickr-SoundNet.
Towards Open-Vocabulary Audio-Visual Event Localization: This work formally defines the Open-Vocabulary Audio-Visual Event Localization (OV-AVEL) task, constructs the OV-AVEBench benchmark containing 24,800 videos across 67 event categories, and proposes two baselines (training-free and fine-tuning) based on ImageBind. Fine-tuning with just a single-layer temporal Transformer achieves an average performance of 57.8%.
Contextual AD Narration with Interleaved Multimodal Sequence: A unified framework named Uni-AD is proposed. It takes interleaved multimodal sequences (video features + text + character bank + context) as input. By aligning features through a visual mapping network, identifying main characters via a character-refinement module, and enhancing contextual consistency with a contrastive loss, it achieves SOTA performance on MAD-eval-Named.
DistinctAD: Distinctive Audio Description Generation in Contexts: Generates distinctive audio descriptions (AD) in contexts to avoid generating generic and featureless descriptions by employing contrastive learning to encourage differences from preceding and succeeding ADs.
EMoVA: Empowering Language Models to See, Hear and Speak with Vivid Emotions: EMoVA is proposed as the first end-to-end omni-modal LLM that achieves visual understanding, speech recognition, and emotion-controllable speech synthesis simultaneously through a semantic-acoustic decoupled speech tokenizer, outperforming GPT-4o on vision-language benchmarks and achieving a 2.9% WER in speech recognition.

Browse all 21 Audio & Speech papers →

🔎 AIGC Detection (3)¶

ProAPO: Progressively Automatic Prompt Optimization for Visual Classification: Proposes ProAPO, a progressively automatic prompt optimization method based on evolutionary algorithms. With only one-shot supervision and zero human intervention, it progressively optimizes from task-level templates to category-level descriptions to address hallucination and lack of discriminativeness in LLM-generated descriptions, outperforming existing text prompting methods on 13 datasets.
SGC-Net: Stratified Granular Comparison Network for Open-Vocabulary HOI Detection: This paper proposes the Stratified Granular Comparison Network (SGC-Net), which aggregates multi-layer CLIP visual features via a Granularity-Aware Alignment (GSA) module and recursively generates discriminative descriptions using an LLM within a Hierarchical Group Comparison (HGC) module. This addresses the issues of insufficient feature granularity and semantic confusion in open-vocabulary HOI detection.
Enhancing Few-Shot Class-Incremental Learning via Training-Free Bi-Level Modality Calibration: This paper proposes the BiMC (Bi-level Modality Calibration) framework based on a frozen CLIP model. By leveraging intra-modal calibration (combining fine-grained class descriptions generated by LLMs with visual prototypes) and inter-modal calibration (fusing pre-trained language knowledge with task-specific visual priors), BiMC achieves state-of-the-art FSCIL performance without any parameter training, outperforming the best baseline by 4.25% on CIFAR-100.

🧊 3D Reconstruction & Rendering (121)¶

UnCommon Objects in 3D: Meta introduces uCO3D—currently the largest public object-centric 3D dataset, containing high-resolution videos of over 1,000 object categories with complete 360° 3D annotations (camera poses, depth maps, point clouds, 3D Gaussian Splatting reconstructions, and text descriptions). Training on this dataset yields significantly better performance on multiple 3D learning tasks compared to MVImgNet and CO3Dv2.
3D-HGS: 3D Half-Gaussian Splatting: This work proposes the 3D Half-Gaussian (3D-HGS) reconstruction kernel, which splits a 3D Gaussian into two halves using a cutting plane, each having independent opacity. Acting as a plug-and-play reconstruction kernel to replace standard Gaussian kernels, it significantly enhances rendering quality at shape and color discontinuities without sacrificing rendering speed, outperforming all SOTA methods on Mip-NeRF360, Tanks & Temples, and Deep Blending.
3D-SLNR: A Super Lightweight Neural Representation for Large-scale 3D Mapping: This paper proposes 3D-SLNR, a super lightweight neural 3D representation. It defines the global Signed Distance Function (SDF) based on a collection of band-limited local SDFs anchored on support points of a point cloud. Each local SDF is parameterized by a single shared tiny MLP (without latent feature vectors). The output of the MLP is modulated by learnable geometric attributes (position, rotation, and scale) to adapt to complex geometries in different regions. Combined with a parallel query algorithm and a prune-and-expand strategy, it achieves SOTA reconstruction quality with less than 1/5 of the memory footprint of previous methods.
3D Convex Splatting: Radiance Field Rendering with 3D Smooth Convexes: Replaces Gaussian primitives with 3D smooth convex primitives for radiance field rendering. By defining convex hulls using point sets + LogSumExp smoothing + custom CUDA rasterizer, this method outperforms 3DGS on T&T and Deep Blending using fewer primitives.
3D Gaussian Inpainting with Depth-Guided Cross-View Consistency: This paper proposes 3DGIC, a framework that achieves object removal and inpainting in 3D Gaussian Splatting scenes through depth-guided cross-view consistent inpainting. By leveraging rendered depth maps, it projects background pixels visible from other views onto the masked region to refine the inpainting mask. Then, 2D inpainting results from a reference view are projected onto 3D space to constrain cross-view consistency for other views. The proposed method outperforms existing approaches in FID and LPIPS on the SPIn-NeRF dataset.
3D Student Splatting and Scooping (SSS): This work proposes SSS (Student Splatting and Scooping), advancing the 3DGS paradigm with three unprecedented innovations: (1) replacing Gaussian distributions with Student-t distributions as mixture components (with learnable tail thickness that varies continuously from Cauchy to Gaussian); (2) introducing negative density components (scooping by subtracting color) to extend the formulation to non-monotonic mixture models; (3) employing SGHMC sampling instead of SGD to decouple parameter optimization. SSS achieves state-of-the-art results in 6 out of 9 metrics across Mip-NeRF360, T&T, and Deep Blending, demonstrating extreme parameter efficiency by matching or exceeding 3DGS using only 18% of the component count.
3DEnhancer: Consistent Multi-View Diffusion for 3D Enhancement: Proposed a 3D enhancement framework based on a multi-view latent diffusion model. By incorporating a pose-aware encoder, multi-view row attention, and adjacent-view epipolar aggregation modules, it significantly enhances the texture quality of low-quality 3D generation results while maintaining cross-view consistency.

3DGUT: Enabling Distorted Cameras and Secondary Rays in Gaussian Splatting

4DEquine: Disentangling Motion and Appearance for 4D Equine Reconstruction from Monocular Video: Disentangles 4D equine reconstruction from monocular video into motion estimation (AniMoFormer spatio-temporal Transformer) and appearance reconstruction (EquineGS single-image feed-forward 3DGS). Leveraging the VAREN parametric model and two large-scale synthetic datasets, it achieves SOTA geometry + appearance reconstruction results on real-world data and generalizes zero-shot to donkeys and zebras.
4DGC: Rate-Aware 4D Gaussian Compression for Efficient Streamable Free-Viewpoint Video: This paper proposes 4DGC, a rate-distortion-aware 4D Gaussian compression framework. By adopting motion-aware dynamic Gaussian modeling (multi-resolution motion grids + sparse compensatory Gaussians) and end-to-end compression (differentiable quantization + implicit entropy model), 4DGC achieves 16× compression over 3DGStream without sacrificing rendering quality.

Browse all 121 3D Reconstruction & Rendering papers →

🧊 3D Generation & Editing (41)¶

Eval3D: Interpretable and Fine-grained Evaluation for 3D Generation: This paper proposes Eval3D, a fine-grained and interpretable evaluation tool for 3D generation quality. The core idea is to utilize various foundation models and tools as probes to detect inconsistencies in the semantic, geometric, structural, and text-alignment aspects of generated 3D assets. This achieves pixel-precise measurements and 3D spatial feedback, aligning more closely with human judgment than existing metrics.
Gen3DEval: Using vLLMs for Automatic Evaluation of Generated 3D Objects: This paper proposes Gen3DEval, a text-to-3D generation quality evaluation framework based on fine-tuned vLLMs. By fine-tuning the Llama3 model on synthetic and human-annotated data, it achieves automatic evaluation of 3D object appearance, surface quality, and text consistency, significantly outperforming general models like GPT-4o in human preference alignment.

CADCrafter: Generating Computer-Aided Design Models from Unconstrained Images

CADDreamer: CAD Object Generation from Single-view Images: This paper proposes CADDreamer, which directly generates CAD models with compact B-rep representations, clear structures, and sharp edges from a single RGB image. Utilizing a semantic-enhanced multi-view diffusion model and a geometric-topological extraction module, it supports five primitive types: planes, cylinders, cones, spheres, and tori.
Instruct-4DGS: Efficient Dynamic Scene Editing via 4D Gaussian-based Static-Dynamic Separation: Instruct-4DGS is proposed, leveraging the inherent separability of static 3D Gaussians and Hexplane deformation fields in 4D Gaussian Splatting (4DGS) to achieve efficient dynamic scene editing by focusing solely on editing static canonical Gaussians. Temporal alignment is refined via Coherent-IP2P-driven score distillation to eliminate motion artifacts, reducing the editing time by more than half while requiring only a single GPU.
ESCAPE: Equivariant Shape Completion via Anchor Point Encoding: ESCAPE proposes a rotation-equivariant point cloud completion method based on anchor distance encoding. By representing point clouds as distance matrices to high-curvature anchor points, the Transformer is enabled to predict the complete shape within a rotation-invariant distance space, and coordinates are subsequently recovered via optimization. This approach significantly outperforms existing methods under arbitrary input rotations (slashing CD-L1 on the PCN dataset from 26.65 to 10.58).
FreeScene: Mixed Graph Diffusion for 3D Scene Synthesis from Free Prompts: FreeScene proposes a user-friendly indoor scene synthesis framework. It utilizes a VLM-driven Graph Designer to convert free-form text/image inputs into scene graphs, and then uses a Mixed Graph Diffusion Transformer (MG-DiT) to perform graph-aware denoising in a hybrid continuous-discrete space. It unifiedly supports multiple tasks such as text-to-scene and graph-to-scene, outperforming existing methods in both generation quality and controllability.
FruitNinja: 3D Object Interior Texture Generation with Gaussian Splatting: FruitNinja proposes the first method for generating internal textures for 3DGS objects. Through progressive cross-section inpainting, voxel smoothing, and the OpaqueAtom GS strategy, it achieves real-time rendering after cutting without additional optimization, significantly outperforming baselines in semantic alignment and texture consistency.
GenPC: Zero-shot Point Cloud Completion via 3D Generative Priors: This paper proposes GenPC, a zero-shot point cloud completion framework. It uses a Depth Prompting module to convert partial point clouds into depth maps to generate RGB images as input for Image-to-3D models. Then, a Geometric Preserving Fusion module aligns and fuses the generated 3D shape with the original point cloud, achieving faster and better real-world scan completion compared to SDS-based optimization methods.
GenVDM: Generating Vector Displacement Maps From a Single Image: The first method to generate Vector Displacement Maps (VDMs) from a single image is proposed. By fine-tuning Zero123++ to generate multi-view normal maps, using neural SDF to reconstruct meshes, and then parameterizing them into VDM images using neural deformation fields, the authors construct the first academic VDM dataset. This provides 3D artists with the ability to generate customized geometric detail stamps on-demand.

Browse all 41 3D Generation & Editing papers →

🧊 3D Representation & Scene Understanding (47)¶

3D-LLaVA: Towards Generalist 3D LMMs with Omni Superpoint Transformer: This work proposes 3D-LLaVA, a general-purpose 3D Large Multimodal Model (LMM) with a minimalist architecture. The core is the Omni Superpoint Transformer (OST) acting as a versatile visual connector. It simultaneously serves as a visual feature selector, a visual prompt encoder, and a segmentation mask decoder. Using only point cloud inputs, it fully achieves state-of-the-art (SOTA) performance across five benchmarks, including ScanQA (92.6 CiDEr) and ScanRefer (43.3 mIoU).
4Deform: Neural Surface Deformation for Robust Shape Interpolation: This paper proposes the 4Deform framework, which achieves robust shape interpolation based on neural implicit representation and continuous velocity field learning. By linking the implicit field and the velocity field via a modified level-set equation, it achieves SOTA performance for the first time across scenarios involving noise, partial observations, topological changes, and non-isometric deformations, while supporting temporal super-resolution of real-world Kinect point cloud sequences.
A2Z-10M+: Geometric Deep Learning with A-to-Z BRep Annotations for AI-Assisted CAD Modeling and Reverse Engineering: The A2Z dataset is constructed containing over 1 million complex CAD models and more than 10 million multimodal annotations (high-resolution 3D scans, freehand 3D sketches, text descriptions, and BRep topology labels), representing the largest dataset for CAD reverse engineering to date. Based on this, foundation models for BRep boundary and corner detection are trained.
Any3DIS: Class-Agnostic 3D Instance Segmentation by 2D Mask Tracking: Any3DIS is proposed, which replaces traditional unsupervised merging strategies with 3D-aware 2D mask tracking (utilizing SAM-2 to track the 2D segmentations of each superpoint across multiple frames) and optimizes 3D proposals using dynamic programming. It achieves state-of-the-art (SOTA) results in class-agnostic, open-vocabulary, and open-ended 3D instance segmentation tasks on ScanNet200 and ScanNet++.
ASHiTA: Automatic Scene-grounded Hierarchical Task Analysis: The first framework, ASHiTA, is proposed to automatically decompose high-level tasks into hierarchies of scene-grounded subtasks. By alternating LLM-assisted hierarchical task analysis with task-driven 3D scene graph construction based on the Information Bottleneck principle, joint reasoning of the task hierarchy and scene representation is achieved.
BFANet: Revisiting 3D Semantic Segmentation with Boundary Feature Analysis: Revisiting 3D semantic segmentation from the perspective of error analysis, this study classifies segmentation errors into four categories (region classification, displacement, merge, and false response) and designs corresponding evaluation metrics. It proposes BFANet, which enhances boundary awareness through a boundary-semantic decoupling module and real-time boundary pseudo-label computation, achieving 36.0 mIoU on the ScanNet200 test set (the highest score without utilizing auxiliary data during training).
Category-Agnostic Neural Object Rigging: Proposes CANOR (Category-Agnostic Neural Object Rigging), which automatically discovers low-dimensional pose spaces of deformable objects in a completely category-agnostic, data-driven manner by encoding deformable 4D objects into a sparse set of spatially localized blobs and instance-aware feature volumes, enabling intuitive pose manipulation.
COB-GS: Clear Object Boundaries in 3DGS Segmentation Based on Boundary-Adaptive Gaussian Splitting: Proposes COB-GS, a boundary-adaptive Gaussian splitting technique driven by semantic gradient statistics. It jointly optimizes semantic information and visual texture to resolve the blurred object boundary issue in 3DGS segmentation, achieving clear object boundary segmentation while preserving visual quality.
CrossOver: 3D Scene Cross-Modal Alignment: The CrossOver framework is proposed to learn a unified scene-level cross-modal embedding space for RGB images, point clouds, CAD models, floor plans, and textual descriptions without requiring complete modal pairing. It utilizes dimensionality-specific encoders and a three-stage training pipeline to support flexible cross-modal retrieval and localization.
Dr. Splat: Directly Referring 3D Gaussian Splatting via Direct Language Embedding Registration: This paper proposes Dr. Splat, which bypasses the rendering process and directly registers language-aligned CLIP embeddings onto 3D Gaussians. Combined with Product Quantization (PQ) pre-trained on large-scale image data, it achieves a 6.25% embedding compression rate. Without requiring any scene-specific optimization (~10 minutes vs. 1–24 hours in prior work), it significantly outperforms existing methods in open-vocabulary 3D semantic segmentation, 3D object grounding, and 3D object selection.

Browse all 47 3D Representation & Scene Understanding papers →

🧊 Depth, Normals & Stereo (23)¶

Blurry-Edges: Photon-Limited Depth Estimation from Defocused Boundaries: This paper proposes a depth estimation method based on a novel image patch representation termed Blurry-Edges. By modeling the smoothness of defocused boundaries, it achieves robust depth estimation under extremely low-light (photon-limited) conditions from a pair of images with different defocus levels, improving noise robustness by over 4 times compared to existing DfD methods.
Consistency-aware Self-Training for Iterative-based Stereo Matching: This paper proposes the first consistency-aware self-training framework (CST-Stereo) for iterative-based stereo matching. It evaluates pseudo-label reliability through multi-resolution prediction consistency filtering and iterative prediction consistency filtering, and combines them with a soft-weighted loss to leverage unlabeled real-world data effectively, thereby improving model performance and generalization.
DEFOM-Stereo: Depth Foundation Model Based Stereo Matching: Integrates the monocular depth foundation model (Depth Anything V2) into the recurrent stereo matching framework RAFT-Stereo. By incorporating combined feature encoders and a scale update module, this approach achieves state-of-the-art stereo matching performance across multiple benchmarks while preserving strong generalization capabilities.
Depth Any Camera: Zero-Shot Metric Depth Estimation from Any Camera: The Depth Any Camera (DAC) framework is proposed, which enables zero-shot metric depth estimation generalizing to fisheye and 360° cameras with training restricted solely to perspective images. By utilizing ERP unified representation, pitch-aware transformation, and FoV alignment, DAC improves $\delta_1$ accuracy by up to 50% on wide-FoV datasets.
DepthCrafter: Generating Consistent Long Depth Sequences for Open-world Videos: Leverages a pretrained video diffusion model (SVD) for video depth estimation. Through a three-stage training strategy, it realizes temporally consistent depth sequence generation of variable lengths (up to 110 frames). By designing a segment-based inference strategy, it supports extremely long videos, comprehensively outperforming existing methods in zero-shot settings.
Dual Exposure Stereo for Extended Dynamic Range 3D Imaging: This paper proposes a dual-exposure stereo method (Dual-Exposure Stereo) that extends the effective dynamic range by automatically controlling the dual-exposure parameters of stereo cameras, and designs a motion-aware dual-exposure depth estimation network to achieve robust 3D imaging in wide dynamic range scenes.
Dual Exposure Stereo for Extended Dynamic Range 3D Imaging: Proposes a Dual-Exposure Stereo method that utilizes Automatic Dual-Exposure Control (ADEC) to apply different exposures in alternating frames, combined with a motion-aware dual-exposure feature fusion network for disparity estimation. This extends the effective dynamic range of stereo cameras to 160% and achieves robust 3D imaging under extreme lighting conditions.
Efficient Depth Estimation for Unstable Stereo Camera Systems on AR Glasses: Proposed two models, MultiHeadDepth and HomoDepth. They optimize the latency bottlenecks of the cost volume and preprocessing in stereo depth estimation using a hardware-friendly multi-head cost volume (approximating cosine similarity via LayerNorm + dot product, combined with group-wise pointwise convolutions) and a homography estimation network with 2D Rectified Positional Encoding (RPE), respectively. In AR glasses scenarios, this improves accuracy by 11.8-30.3% while reducing end-to-end latency by 44.5%.
Feature-Preserving Mesh Decimation for Normal Integration: The classic quadric error metric (QEM) is derived in screen space with normal maps as input. Combined with optimal Delaunay triangulation, this achieves anisotropic mesh decimation that preserves sub-millimeter accuracy even at 90%+ compression rates, accelerating high-resolution normal integration from hours to minutes.
FoundationStereo: Zero-Shot Stereo Matching: Presents FoundationStereo, a large-scale foundation model for stereo depth estimation. By leveraging a million-scale high-fidelity synthetic dataset, fusing monocular depth priors via a Side-Tuning Adapter, and adopting a hybrid cost volume filtering mechanism (incorporating Axial-Planar Convolution and Disparity Transformer), this method achieves strong zero-shot generalization performance without requiring target-domain fine-tuning.

Browse all 23 Depth, Normals & Stereo papers →

🧊 Geometric Matching & Localization (38)¶

4DTAM: Non-Rigid Tracking and Mapping via Dynamic Surface Gaussians: This paper presents the first 4D tracking and mapping method (4DTAM) based on differentiable rendering and 2D Gaussian surface primitives. By jointly optimizing camera poses, scene geometry, appearance, and dynamic deformation fields, 4DTAM achieves real-time reconstruction of non-rigid dynamic scenes from monocular RGB-D video streams. It also releases a novel synthetic 4D dataset, Sim4D, for evaluation.
ActiveGAMER: Active GAussian Mapping through Efficient Rendering: ActiveGAMER is proposed, representing the first attempt to utilize 3D Gaussian Splatting for active mapping. By efficiently selecting the optimal next-best-view via a rendering-based information gain module, integrated with a coarse-to-fine exploration strategy, post-refinement, and a global-local keyframe policy, ActiveGAMER significantly outperforms NeRF-based methods in both geometric accuracy and rendering fidelity on the Replica and MP3D datasets.
AerialMegaDepth: Learning Aerial-Ground Reconstruction and View Synthesis: This paper proposes the AerialMegaDepth dataset generation framework. By co-registering pseudo-synthetic aerial renders from Google Earth and real ground-level images from MegaDepth into a unified coordinate system, it constructs a large-scale training dataset of 132k mixed-altitude images. Fine-tuning DUSt3R on this dataset improves the camera rotation estimation accuracy for aerial-ground pairs from 5% to 56%, while also significantly enhancing the quality of novel view synthesis.
CMMLoc: Advancing Text-to-PointCloud Localization with Cauchy-Mixture-Model Based Framework: Proposes CMMLoc, an uncertainty-aware text-to-point cloud localization framework based on the Cauchy Mixture Model (CMM). By modeling the coarse retrieval stage as a partially relevant retrieval problem and introducing a CMM Transformer and a cardinal direction integration module, it achieves SOTA performance on the KITTI360Pose dataset.
ColabSfM: Collaborative Structure-from-Motion by Point Cloud Registration: The ColabSfM paradigm is proposed, which merges distributed SfM reconstruction results via 3D point cloud registration (rather than visual descriptor matching). In addition, a dedicated SfM registration dataset generation pipeline and an improved registration model, RefineRoITr, are designed.
CoMatcher: Multi-View Collaborative Feature Matching: Proposes CoMatcher, a multi-view collaborative feature matcher that shifts from the independent two-view matching paradigm to a 1-to-N collaborative matching paradigm, leveraging contextual cues from complementary views and cross-view projective consistency constraints to improve matching reliability in complex scenes.
Cross-View Completion Models are Zero-shot Correspondence Estimators: Reveals that the cross-attention maps in cross-view completion (CVC) models inherently learn precise dense correspondence, and proposes ZeroCo to leverage this finding in zero-shot matching and learning-based geometric matching, significantly outperforming conventional methods based on encoder/decoder features.
Denoising Functional Maps: Diffusion Models for Shape Correspondence: This paper proposes DenoisFM, the first method to apply denoising diffusion models to directly predict functional maps between shapes. It reduces learning complexity through template matching and introduces an unsupervised approach to resolve the sign ambiguity of Laplace eigenvectors, achieving competitive performance on human and animal shape matching.
Dense-SfM: Structure from Motion with Dense Consistent Matching: Proposed the Dense-SfM framework, which resolves the fragmented track issue of dense matches through Gaussian Splatting-based track expansion, and combines a Transformer and Gaussian Process-based multi-view kernelized match refinement module to achieve high-precision dense SfM reconstruction.
Doppelgangers++: Improved Visual Disambiguation with Geometric 3D Features: This paper proposes Doppelgangers++, which significantly improves the precision and generalization of doppelganger (visually ambiguous image pair) detection by introducing diverse daily scenes training data from VisymScenes and training a Transformer classifier using 3D-aware features from the multi-layer decoder of MASt3R. It seamlessly integrates into COLMAP and MASt3R-SfM pipelines to improve 3D reconstruction quality in scenes with repetitive structures.

Browse all 38 Geometric Matching & Localization papers →

🎯 Object Detection (38)¶

AA-CLIP: Enhancing Zero-Shot Anomaly Detection via Anomaly-Aware CLIP: Proposed AA-CLIP, which enhances anomaly discriminability while preserving the generalization ability of CLIP through a two-stage training strategy (first adapting the text encoder to establish anomaly-aware anchors, then aligning patch-level visual features). It achieves SOTA zero-shot anomaly detection performance across multiple industrial and medical datasets with minimal training samples.
T2ICount: Enhancing Cross-modal Understanding for Zero-Shot Counting: This paper proposes T2ICount, which leverages one-step denoising features from pre-trained text-to-image diffusion models for zero-shot object counting. It addresses the lack of text sensitivity in one-step denoising through a Hierarchical Semantic Correction Module (HSCM) and a Representational Regional Coherence loss ($\mathcal{L}_{RRC}$).
Distribution Prototype Diffusion Learning for Open-set Supervised Anomaly Detection: The DPDL method is proposed to learn multi-Gaussian distribution prototypes and map normal samples to the prototype space via diffusion using the Schrödinger Bridge (while concurrently pushing away anomalous samples). Combined with dispersion feature learning on hyperspherical space to enhance generalization, this method achieves state-of-the-art (SOTA) performance on 9 public anomaly detection datasets (e.g., outperforming AHL by 5.0% on AITEX and 8.7% on ELPV).
DEIM: DETR with Improved Matching for Fast Convergence: This paper accelerates DETR training convergence through two simple improvements: Dense O2O (increasing targets per image via data augmentation to achieve dense one-to-one matching) and MAL (replacing VFL to better optimize low-quality matches). It cuts the training epochs in half while boosting performance (COCO AP 56.5 with D-FINE-X).
Efficient Event-Based Object Detection: A Hybrid Neural Network with Spatial and Temporal Attention: This work proposes the first hybrid SNN-ANN object detection model targeting large-scale benchmarks. An Attention-Squeeze Bridging (ASAB) block is designed to convert sparse spike representations from the SNN into dense features for the ANN via spatio-temporal attention. With only 6.6M parameters, it significantly outperforms SNN methods and approaches the accuracy of ANN/RNN methods on the Gen1/Gen4 datasets, while the SNN component can be deployed on the Intel Loihi 2 neuromorphic chip for low-power inference.
Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models: The first MLLM dedicated to zero-shot anomaly detection and reasoning (Anomaly-OV). It generates anomaly saliency maps through a Look-Twice Feature Matching mechanism coupled with a visual token selector to focus on suspicious regions, achieving SOTA zero-shot anomaly detection with an average AUROC of 88.6% across 9 benchmarks.
Interpreting Object-level Foundation Models via Visual Precision Search: Addressing the interpretability of object-level foundation models like Grounding DINO and Florence-2, this paper proposes Visual Precision Search (VPS). By combining superpixel sparsification with greedy search guided by submodular functions, VPS accurately localizes critical decision subregions. It outperforms the state-of-the-art D-RISE method in fidelity metrics (Insertion) by 23.7%, 20.1%, and 31.6% on MS COCO, RefCOCO, and LVIS, respectively.
Multiple Object Tracking as ID Prediction: This paper proposes MOTIP, which reformulates the object association problem in multiple object tracking (MOT) as an in-context ID prediction task. Given historical trajectories with ID embeddings, a standard Transformer decoder directly predicts the ID labels of the current detections without relying on heuristic matching algorithms, achieving a HOTA of 69.6 on DanceTrack and significantly outperforming the previous SOTA CO-MOT (65.3).
One-for-More: Continual Diffusion Model for Anomaly Detection: The CDAD framework is proposed to achieve stable continual learning for diffusion models via gradient projection. Supported by iterative SVD (iSVD), the memory consumption is reduced from 157GB to 17GB. Additionally, an anomaly-masked network is designed to enhance the conditioning mechanism, achieving first place in 17 out of 18 settings across MVTec and VisA.
MI-DETR: An Object Detection Model with Multi-time Inquiries Mechanism: MI-DETR proposes a parallel Multi-time Inquiries (MI) mechanism to replace the traditional cascaded decoder architecture of DETR. This allows object queries to learn multi-modal information from image features in parallel through multiple parameter-independent inquiry heads. Combined with U-like Feature Interaction (UFI), it achieves 52.7 AP on COCO with a ResNet-50 backbone, outperforming all existing DETR variants.

Browse all 38 Object Detection papers →

✂️ Segmentation (96)¶

OverLoCK: An Overview-first-Look-Closely-next ConvNet with Context-Mixing Dynamic Kernels: OverLoCK is proposed, which is the first pure convolutional backbone network that explicitly incorporates a top-down attention mechanism. Through a deep-stage decomposition strategy (DDS) and context-mixing dynamic convolution (ContMix), it surpasses ConvNeXt-B on ImageNet-1K using only 1/3 of the FLOPs, achieving comprehensive leadership in detection and segmentation tasks.
Frequency Dynamic Convolution for Dense Image Prediction: FDConv redesigns dynamic convolution from a frequency domain perspective. By leveraging Fourier Disjoint Weights (FDW), it constructs frequency-diverse convolutional kernels without increasing parameters. Combined with Kernel Spatial Modulation (KSM) and Frequency Band Modulation (FBM) for fine-grained frequency adaptation, FDConv outperforms existing dynamic convolution methods requiring 65-90M extra parameters while only introducing 3.6M parameters.
Your ViT is Secretly an Image Segmentation Model: This paper proposes the Encoder-only Mask Transformer (EoMT), demonstrating that with large-scale pre-training and sufficiently large models, plain ViT can achieve high-quality image segmentation without task-specific components such as CNN adapters, pixel decoders, and Transformer decoders, while being up to 4x faster.
GLUS: Global-Local Reasoning Unified into A Single Large Language Model for Video Segmentation: This paper proposes the GLUS framework, which unifies global understanding and local temporal consistency into a single MLLM through a frame partition strategy of "context frames (global reasoning) + query frames (local tracking)". Combined with an end-to-end trained VOS memory bank module, it significantly outperforms all MLLM-based methods on MeViS (J&F 51.3%).
ResCLIP: Residual Attention for Training-free Dense Vision-language Inference: Discovers that cross-correlation self-attention in the intermediate layers of CLIP possesses localization properties, and proposes two plug-and-play modules: Residual Cross-correlation Self-attention (RCS) and Semantic Feedback Refinement (SFR), significantly improving CLIP's dense inference capabilities in open-vocabulary semantic segmentation.
Golden Cudgel Network for Real-Time Semantic Segmentation: This paper proposes GCNet, featuring the Golden Cudgel Block (GCBlock) as its core. It self-expands during training (multi-convolution, multi-path) to enhance learning capacity, and self-contracts during inference (reparameterized into a single $3\times3$ convolution) to accelerate speed. This yields a "self-distillation" paradigm without requiring an external teacher model, outperforming existing real-time segmentation models on Cityscapes with 77.3% mIoU at 193.3 FPS.
ROS-SAM: High-Quality Interactive Segmentation for Remote Sensing Moving Object: ROS-SAM adapts SAM to the high-quality interactive segmentation task of moving objects in remote sensing videos by fine-tuning the encoder via LoRA, improving the HQ decoder, and redesigning the data pipeline. This achieves a 13% IoU improvement and demonstrates strong zero-shot generalization capabilities.
MambaOut: Do We Really Need Mamba for Vision?: Through conceptual analysis, this paper points out that the SSM mechanism in Mamba is suited for long-sequence + autoregressive tasks, neither of which is satisfied by ImageNet image classification. Consequently, the authors construct the MambaOut series (a pure Gated CNN) by removing SSM. MambaOut completely outperforms all state-of-the-art vision Mamba models on image classification, thoroughly demonstrating that SSM is unnecessary for visual classification.
MambaVision: A Hybrid Mamba-Transformer Vision Backbone: NVIDIA proposes MambaVision, the first systematic study of hybrid Mamba-Transformer formulations for vision backbones. By redesigning the MambaVision Mixer and adding self-attention in the final blocks, it addresses the limitation of SSMs in capturing global context. It achieves a new Pareto front for accuracy-throughput on ImageNet-1K, while also outperforming comparable competitors in downstream detection and segmentation tasks.
MammAlps: A Multi-view Video Behavior Monitoring Dataset of Wild Mammals in the Swiss Alps: This paper proposes MammAlps—a multimodal, multi-view dataset for monitoring the behavior of wild mammals in the Swiss National Park (8.5 hours of dense annotation, 5 species, 11 activities + 19 actions), along with two benchmark tasks: multimodal species + hierarchical behavior recognition (B1) and the first multi-view long-term event understanding (B2), filling the gap in wildlife video behavior analysis regarding hierarchical behavior annotation, multimodality, and multi-view.

Browse all 96 Segmentation papers →

🖼️ Image Restoration (43)¶

HVI: A New Color Space for Low-light Image Enhancement: This paper proposes a new color space HVI (Horizontal/Vertical-Intensity), which eliminates red artifacts through polarized HS mapping, compresses black artifacts in dark regions using a learnable intensity component, and outperforms existing low-light enhancement state-of-the-art methods across 10 datasets in combination with the decoupling network CIDNet.
Iterative Predictor-Critic Code Decoding for Real-World Image Dehazing: IPC-Dehaze proposes an iterative Predictor-Critic decoding framework based on a VQGAN codebook prior. By utilizing a Code-Critic to evaluate the inter-relations among codebook sequences to determine which codes should be retained or resampled, the framework achieves progressive, easy-to-hard dehazing from clear regions to dense haze regions, significantly surpassing state-of-the-art methods in real-world scenarios.
PolarFree: Polarization-based Reflection-Free Imaging: A large-scale RGB-polarization dataset, PolaRGB, consisting of 6,500 pairs was constructed. A two-stage network, PolarFree, was proposed, which first employs a conditional diffusion model to generate a reflection-free prior and then utilizes a de-reflection backbone network to separate the transmission layer. This approach outperforms previous methods by approximately 2dB PSNR in polarization-guided reflection removal.
PIDSR: Complementary Polarized Image Demosaicing and Super-Resolution: PIDSR proposes a framework for joint complementary optimization of polarized image demosaicing (PID) and polarization image super-resolution (PISR). Utilizing a two-stage recurrent pipeline (spatial-physical coherent reconstruction + polarization-aware resolution enhancement) and a Stokes-assisted network, it directly reconstructs high-quality high-resolution polarization images from CPFA raw images, significantly reducing errors in DoP and AoP.
Rotation-Equivariant Self-Supervised Method in Image Denoising: This work introduces rotation-equivariant convolutions to self-supervised image denoising for the first time. It rigorously analyzes the impact of up/downsampling operators on equivariance, provides the equivariant error bounds of the complete U-Net architecture, and proposes an adaptive rotation-equivariant network, AdaReNet. Through a learning-based mask fusion module, AdaReNet automatically determines which regions of an image are better suited for the rotation-equivariant network, achieving consistent performance improvements across three classic self-supervised methods: N2N, N2V, and R2R.
Augmenting Perceptual Super-Resolution via Image Quality Predictors: No-reference image quality assessment (NR-IQA) models are leveraged to replace human annotations. By improving perceptual super-resolution quality through weighted sampling and direct optimization, the proposed method outperforms state-of-the-art methods that rely on human feedback, without requiring any human-labeled data.
FiRe: Fixed-points of Restoration Priors for Solving Inverse Problems: This paper proposes the FiRe framework, which formulates explicit image priors using fixed-point theory by composing general-purpose image restoration models (e.g., deblurring, super-resolution, inpainting) with their training degradation operators. This generalizes traditional Plug-and-Play (PnP) beyond denoiser-only priors and supports the ensemble of multiple restoration models, significantly outperforming existing PnP and diffusion-based approaches on various inverse problems.
DarkIR: Robust Low-Light Image Restoration: DarkIR proposes an efficient CNN-based multi-task low-light image restoration method. The encoder uses SpAM+FreMLP (frequency magnitude enhancement) to handle illumination, while the decoder utilizes Di-SpAM (dilated spatial attention) to handle blur. With an asymmetric design, it achieves 27.30dB PSNR on LOLBlur with only 3.31M parameters.
Efficient Visual State Space Model for Image Deblurring: This paper proposes EVSSM, which efficiently captures non-local information by applying alternating geometric transformations (transpose/flip) before unidirectional SSM scanning, and designs an efficient discriminative frequency-domain FFN (EDFFN) to enhance local details. It outperforms existing SSM methods and achieves SOTA on image deblurring tasks with only 1/4 of the computational cost.
MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration: This work proposes MaIR. Its key innovations include a Nested S-shaped Scanning (NSS) strategy that preserves locality via stripe partitioning and continuity via S-shaped paths, and a Sequence Shuffle Attention (SSA) module that intelligently aggregates sequences from different scanning directions using channel-level attention. MaIR achieves state-of-the-art (SOTA) performance across 14 datasets in four major tasks: super-resolution, denoising, deblurring, and dehazing.

Browse all 43 Image Restoration papers →

🛰️ Remote Sensing (11)¶

Meta-Learning Hyperparameters for Parameter Efficient Fine-Tuning: MetaPEFT proposes a meta-learning framework that unifies discrete position selection and continuous scaling factors in PEFT into differentiable modulators. Through bi-level optimization, it automatically searches for the optimal PEFT hyperparameter configuration, achieving SOTA on remote sensing and natural image long-tailed distribution adaptation tasks.
Learning Occlusion-Robust Vision Transformers for Real-Time UAV Tracking: The ORTrack framework is proposed to learn occlusion-robust ViT feature representations through random masking based on spatial Cox processes (imposing mask constraints during training and achieving zero overhead during inference). An adaptive feature distillation method is designed to compress large models into a lightweight student model ORTrack-D, achieving the best balance of state-of-the-art accuracy and real-time speed across several UAV tracking benchmarks.
Dense Dispersed Structured Light for Hyperspectral 3D Imaging of Dynamic Scenes: This paper proposes the Dense Dispersed Structured Light (DDSL) method, which utilizes an inexpensive diffraction grating film (<$20), a stereo RGB camera, and an RGB projector. By designing spectrally multiplexed DDSL patterns, the required number of projection frames is significantly reduced, achieving real-time hyperspectral 3D imaging at 6.6 fps with a spectral resolution of 15.5 nm FWHM and a depth error of 4 mm.
DiSciPLE: Learning Interpretable Programs for Scientific Visual Discovery: The DiSciPLE framework is proposed to automatically synthesize interpretable Python programs for visual data analysis using an LLM-guided evolutionary algorithm. It achieves SOTA on scientific tasks such as population density estimation, reducing error by 35% compared to recent baselines while remaining fully interpretable.
EarthDial: Turning Multi-sensory Earth Observations to Interactive Dialogues: This work proposes EarthDial, a conversational vision-language model tailored for Earth Observation (EO) data. It supports the unified understanding of multispectral (SAR/NIR/infrared), multi-temporal, and multi-resolution remote sensing imagery. Trained on an 11.11 million instruction-tuning dataset, it outperforms existing remote sensing VLMs across 44 downstream datasets.
Hierarchical Dual-Change Collaborative Learning for UAV Scene Change Captioning: This paper proposes a new task named UAV Scene Change Captioning (UAV-SCC) and a novel HDC-CL framework. It models the overlapping and non-overlapping regions of image pairs under moving viewpoints using a Dynamic Adaptive Layout Transformer, enhances viewpoint shift direction awareness via hierarchical cross-modal directional consistency calibration, and constructs a dedicated benchmark dataset.
Joint and Streamwise Distributed MIMO Satellite Communications with Multi-Antenna Ground Users: This paper investigates distributed MIMO downlink communications where multiple LEO satellites jointly serve multi-antenna ground users. Two modes, namely joint transmission and streamwise transmission, are proposed. The former optimizes the precoder using WMMSE iterations to maximize the sum spectral efficiency, while the latter employs a Hungarian algorithm-based stream-satellite association to reduce the fronthaul overhead, achieving a flexible trade-off between performance and the fronthaul signaling load.
MetaSpectra+: A Compact Broadband Metasurface Camera for Snapshot Hyperspectral+ Imaging: Proposed MetaSpectra+, a compact and multi-functional camera based on hybrid metasurface-refractive optics. By utilizing double-layer metasurfaces to independently control dispersion, exposure, and polarization for each channel, it achieves snapshot hyperspectral+HDR or hyperspectral+polarization joint imaging within an ~250nm visible bandwidth, achieving SOTA reconstruction accuracy on benchmark datasets.
MFogHub: Bridging Multi-Regional and Multi-Satellite Data for Global Marine Fog Detection and Forecasting: MFogHub constructs the first multi-regional (15 coastal regions) and multi-satellite (6 geostationary satellites) global marine fog detection and forecasting dataset, containing over 68,000 high-resolution samples and 11,600+ pixel-level annotations. Extensive experiments on 16 baseline models reveal the influence of regional differences and satellite variations on model generalization.
SGFormer: Satellite-Ground Fusion for 3D Semantic Scene Completion: This work introduces satellite imagery to the 3D Semantic Scene Completion (SSC) task for the first time, proposing a dual-branch framework named SGFormer. By utilizing ground-view guided satellite feature correction and adaptive fusion strategies, it effectively addresses the scene incompleteness issue caused by visual occlusions.

Browse all 11 Remote Sensing papers →

🧑 Human Understanding (109)¶

Any6D: Model-free 6D Pose Estimation of Novel Objects: This paper proposes the Any6D framework to estimate the 6D pose and size of novel objects from a single RGB-D anchor image. By combining InstantMesh 3D reconstruction, oriented bounding box coarse alignment, and joint size-pose refinement, Any6D achieves an ADD-S of 98.7% on HO3D, significantly outperforming GEDI's 71.9%.
RGBAvatar: Reduced Gaussian Blendshapes for Online Modeling of Head Avatars: RGBAvatar proposes a "Reduced Gaussian Blendshapes" representation that efficiently represents animatable head avatars using only 20 learnable bases. Combined with batch-parallel rendering and a color initialization strategy, it achieves online real-time (reconstructing while capturing) head avatar reconstruction for the first time.
FSFM: A Generalizable Face Security Foundation Model via Self-Supervised Facial Representation Learning: FSFM proposes the first self-supervised pre-training framework specialized for face security tasks. By employing the CRFR-P facial masking strategy coupled with a dual-task collaborative learning of MIM/ID, it learns "3C" representations of real faces (intra-region consistency, inter-region coherence, and local-to-global correspondence). It surpasses task-specific SOTA methods across three major tasks: Deepfake Detection, Face Anti-Spoofing, and Diffusion Forgery Detection.
VTON 360: High-Fidelity Virtual Try-On from Any Viewing Direction: VTON 360 is proposed, which reformulates 3D virtual try-on as a multi-view consistent 2D virtual try-on expansion problem. By combining pseudo-3D pose representation, multi-view spatial attention, and multi-view CLIP embedding, it achieves high-fidelity virtual try-on from arbitrary viewing directions.
Co-op: Correspondence-based Novel Object Pose Estimation: This paper proposes Co-op, a correspondence-based 6DoF pose estimation framework for novel objects. In the coarse estimation stage, a hybrid representation (patch-level classification + offset regression) is used to estimate the initial pose quickly and accurately with only 42 templates. In the refinement stage, probabilistic flow regression combined with differentiable PnP is utilized for end-to-end optimization, significantly outperforming existing methods on seven core datasets of the BOP Challenge.
RePerformer: Immersive Human-centric Volumetric Videos from Playback to Photoreal Reperformance: RePerformer is proposed, a 3DGS-based volumetric video representation that unifies high-fidelity playback and photorealistic reperformance under novel poses through hierarchical decoupling of motion and appearance Gaussians, Morton coding parameterization, and a semantic-aware alignment module.
UniHOPE: A Unified Approach for Hand-Only and Hand-Object Pose Estimation: This paper proposes UniHOPE, the first framework to unify Hand-Only Pose Estimation (HPE) and Hand-Object Pose Estimation (HOPE). It dynamically controls outputs via an object switcher, eliminates interference from irrelevant object features through grasp-aware feature fusion, and learns occlusion-invariant features using diffusion-based generative de-occlusion combined with multi-level feature enhancement.
StickMotion: Generating 3D Human Motions by Drawing a Stickman: The StickMotion framework is proposed, which uses user-hand-drawn stickman drawings as fine-grained motion control conditions, combined with text descriptions, to achieve global and local 3D human motion generation. A Multi-Condition Module (MCM) is designed to efficiently process condition combinations, saving users 51.5% of their time for expressing creative motion ideas.
Analyzing the Synthetic-to-Real Domain Gap in 3D Hand Pose Estimation: This paper presents the first systematic study of the synthetic-to-real domain gap in 3D hand pose estimation. By designing a controllable data synthesis pipeline, the authors decompose and analyze the impacts of four key factors: forearms, spectral statistics, pose distribution, and object occlusion. The study demonstrates that with proper integration of these factors, purely synthetic data can achieve accuracy on par with real data.
FRESA: Feedforward Reconstruction of Personalized Skinned Avatars from Few Images: This paper proposes FRESA, which learns a general clothed human prior to jointly infer personalized canonical shape, skinning weights, and pose-dependent deformations in a feedforward manner (18 seconds) from a few images. This achieves high-quality animatable 3D human avatar reconstruction with zero-shot generalization to mobile phone photos.

Browse all 109 Human Understanding papers →

📹 Video Understanding (70)¶

T*: Re-thinking Temporal Search for Long-Form Video Understanding: Proposes T*, a lightweight temporal search framework that reformulates expensive temporal search as a spatial search problem. It iteratively localizes keyframes in both temporal and spatial dimensions via an adaptive zooming mechanism. Combined with LV-Haystack, the first large-scale benchmark for long-form video keyframe search, it significantly improves the performance of existing VLMs in long-form video understanding.
Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking: A significant level of layer redundancy (feature saturation) is identified in the deep layers of lightweight ViT trackers. SGLATrack, a similarity-guided layer-adaptive approach, is proposed to dynamically disable redundant layers and retain only a single optimal layer, achieving real-time UAV tracking at 225 FPS on GPU.
DPFlow: Adaptive Optical Flow Estimation with a Dual-Pyramid Framework: This paper proposes DPFlow, a dual-pyramid recurrent encoder combining an image pyramid and a feature pyramid with a fully-convolutional Cross-Gated Unit (CGU). Trained only on standard resolutions, DPFlow adaptively generalizes to 8K resolution inputs, achieving state-of-the-art (SOTA) performance on Sintel, KITTI, and Spring benchmarks. Furthermore, the paper introduces Kubric-NK, a multi-resolution optical flow evaluation dataset, supporting quantitative high-resolution evaluation for the first time.
MUST: The First Dataset and Unified Framework for Multispectral UAV Single Object Tracking: This work proposes the first large-scale multispectral UAV single object tracking dataset, MUST (250 sequences, 43K frames, 8 spectral bands), and designs a unified framework named UNTrack to fuse spectral, spatial, and temporal features, achieving efficient and robust tracking via an asymmetric Transformer and a spectral prompt encoder.
Video Streaming Thinking: VideoLLMs Can Watch and Think Simultaneously: The Video Streaming Thinking (VST) paradigm is proposed to alternate between "watching" and "thinking" during video playback. The model generates intermediate reasoning chains while receiving video frames, amortizing CoT computation into the pre-query phase. This achieves a state-of-the-art (SOTA) score of 79.5% on StreamingBench while maintaining real-time responsiveness (0.56s QA latency).
QA-TIGER: Question-Aware Gaussian Experts for Audio-Visual Question Answering: QA-TIGER is proposed, a framework that models video temporal sequence using continuous adaptive weighting with a Mixture of Gaussian Experts (MoE), and injects question information early in the encoding process to achieve progressive semantic refinement, reaching SOTA on multiple AVQA benchmarks.
STOP: Integrated Spatial-Temporal Dynamic Prompting for Video Understanding: Proposes STOP, an integrated spatial-temporal dynamic prompting method for video understanding. It adaptively highlights discriminative regions via intra-frame spatial prompts and dynamically inserts prompt tokens between frames with high temporal variations via inter-frame temporal prompts, guiding the frozen CLIP model to focus on key spatial-temporal locations.
ViTED: Video Temporal Evidence Distillation: ViTED proposes a framework that automatically generates temporal grounding chain of evidence, unifying evidence collection, temporal grounding, and question-answering reasoning into a single video-language model, enhancing complex video QA capabilities through evidence distillation.
Unbiasing through Textual Descriptions: Mitigating Representation Bias in Video Benchmarks: Ours proposes the UTD method, which leverages VLMs and LLMs to generate video frame textual descriptions to systematically analyze object, temporal, and commonsense biases in video benchmarks, and constructs unbiased test splits to make video understanding evaluation more robust and unbiased.
PAVE: Patching and Adapting Video Large Language Models: PAVE proposes a framework to adapt pre-trained Video LLMs via lightweight "patches," integrating side-channel signals such as audio, 3D cues, and multi-view videos into the base model with only about 0.1% additional parameters and computation, outperforming specialized models on tasks like audio-visual QA and 3D QA.

Browse all 70 Video Understanding papers →

🚗 Autonomous Driving (93)¶

GaussianWorld: Gaussian World Model for Streaming 3D Occupancy Prediction: This paper proposes GaussianWorld, which reformulates 3D occupancy prediction as a 4D occupancy prediction problem conditioned on current sensor inputs. By decomposing scene evolution into three factors — ego-motion alignment, dynamic object motion, and new region completion — the proposed method explicitly models scene changes in the 3D Gaussian space via a world model. Without introducing extra computational overhead, it improves the mIoU of single-frame methods by over 2% on nuScenes.
SceneDiffuser++: City-Scale Traffic Simulation via a Generative World Model: SceneDiffuser++ is proposed, an end-to-end city-scale traffic simulation diffusion model that handles agent spawning and despawning in sparse tensors via soft clipping, achieving over 60 seconds of trip-level traffic simulation with a combined JS divergence of 0.2423 on WOMD-XLMap.
RENO: Real-Time Neural Compression for 3D LiDAR Point Clouds: RENO proposes Sparse Occupancy Codes and a one-time inference strategy, achieving the first real-time neural compression of 3D LiDAR point clouds (10fps@14-bit). With a model size of only 1MB, it outperforms the G-PCC standard by 12.25% in bitrate savings.
SparseAlign: A Fully Sparse Framework for Cooperative Object Detection: SparseAlign proposes the first fully sparse framework for cooperative object detection. By resolving the problems of center feature missing and isolated convolution domains via coordinate-expandable sparse convolution, it outperforms dense BEV-based state-of-the-art methods while reducing communication bandwidth by 98%.
SuperPC: A Single Diffusion Model for Point Cloud Completion, Upsampling, Denoising, and Colorization: SuperPC is proposed as the first framework to unify point cloud completion, upsampling, denoising, and colorization within a single conditional diffusion model. It effectively fuses image and point cloud modalities using a three-level conditioning (TLC) mechanism (raw/local/global) and a spatial mixed fusion (SMF) strategy.
Towards Autonomous Micromobility through Scalable Urban Simulation: This paper proposes URBAN-SIM (a high-performance urban robot learning simulation platform) and URBAN-BENCH (a benchmark with 8 micromobility tasks). By incorporating three core modules—hierarchical urban scene generation, interactive dynamics generation, and asynchronous scene sampling—the framework enables training and evaluation of embodied agents in large-scale, diverse urban environments, serving as a systematic simulation solution for driving autonomous micromobility forward.
DecoupledGaussian: Object-Scene Decoupling for Physics-Based Interaction: Decouples objects from the background in 3DGS scenes to support physical simulations (such as collisions and grasping) while maintaining high-quality rendering.
LightLoc: Learning Outdoor LiDAR Localization at Light Speed: This paper proposes LightLoc, which achieves a 50x acceleration in large-scale outdoor LiDAR localization training (1 hour vs. 2 days) while attaining a state-of-the-art (SOTA) position accuracy of 0.83m. This is achieved via Sample Classification Guidance (SCG) to reduce regression ambiguity in visually similar areas, and Redundant Sample Downsampling (RSD) to discard already well-learned frames.
Segment Anything, Even Occluded: Proposes SAMEO, which adapts EfficientSAM as an amodal segmentation decoder for occluded objects. Combined with a newly constructed 300K-image Amodal-LVIS dataset, it achieves zero-shot amodal segmentation performance on COCOA-cls and D2SA that outperforms supervised methods.
WeatherGen: A Unified Diverse Weather Generator for LiDAR Point Clouds via Spider Mamba Diffusion: This paper proposes WeatherGen, the first unified diffusion generation framework for diverse adverse weather LiDAR data. By preserving the physical structure of LiDAR through the Spider Mamba generator and achieving controllable weather generation via a contrastive learning-based controller, it significantly outperforms physics-based simulation methods in both data fidelity and downstream detection performance.

Browse all 93 Autonomous Driving papers →

🤖 Robotics & Embodied AI (44)¶

MoManipVLA: Transferring Vision-Language-Action Models for General Mobile Manipulation: Proposes MoManipVLA to transfer pre-trained fixed-base VLA models to mobile manipulation scenarios. By jointly planning base movement and manipulator trajectories using a bi-level trajectory optimization (optimizing reachability, smoothness, and collision avoidance), it achieves a 66.1% success rate (+4.2%) on the OVMM benchmark and can be deployed in the real world with only 50 demonstrations.
Phoenix: A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction: This paper proposes the Phoenix framework, which utilizes motion instructions as a bridge to connect the high-level semantic reflection of MLLMs with low-level robotic action correction. By incorporating a dual-process motion adjustment mechanism and a motion-conditioned diffusion policy, Phoenix achieves fine-grained manipulation failure recovery and supports self-improvement through lifelong learning.
Robotic Visual Instruction: Proposes Robotic Visual Instruction (RoVI), a visual instruction paradigm centered on hand-drawn arrows and circles to guide robotic manipulation instead of natural language, and designs the VIEW pipeline to translate 2D visual instructions into 3D action sequences, achieving an 87.5% success rate in real-world environments.
Reasoning in Visual Navigation of End-to-end Trained Agents: A Dynamical Systems Approach: Through large-scale experiments on 262 real-world robot navigation episodes, this work deeply analyzes the emergent reasoning capabilities inside end-to-end RL-trained navigation agents, including a Kalman-filter-like dynamical model, latent memory of scene structures, a finite horizon of planning ability, and value functions associated with long-term planning.
A Data-Centric Revisit of Pre-Trained Vision Models for Robot Learning: Through systematic evaluation, it is found that DINO/iBOT outperforms MAE in robot tasks but suffers performance degradation on non-object-centric (NOC) data due to the loss of object-centric representation capabilities. This paper proposes SlotMIM, which uses a semantic bottleneck (reducing prototype numbers to encourage the emergence of objectness), cross-view consistency regularization, and slot-level contrastive learning. This enables the model to learn object-centric representations from NOC data, outperforming MVP/VC-1 pre-trained on >1M samples using only 241K samples.
Learning Physics-Based Full-Body Human Reaching and Grasping from Brief Walking References: By utilizing only approximately 30 seconds of walking MoCap data and combining transferable movement patterns from walking (shallow network feature alignment) with kinematic-method-generated grasping poses (active data augmentation strategy), this work achieves physically feasible and natural full-body human reach-and-grasp motion generation, achieving a 99.8% grasp success rate in simple scenarios.
Prof. Robot: Differentiable Robot Rendering without Static and Self-Collisions: Prof. Robot is proposed as the first differentiable robot rendering framework incorporating collision constraints. By binding 3D Gaussian points to each link of a robot's URDF model, differentiable rendering is achieved. Concurrently, static collision (with the environment) and self-collision (within the robot itself) constraints are integrated into the optimization, reducing the collision rate from 24% to 0% while maintaining visual fidelity.
CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models: This paper proposes CoT-VLA, which introduces visual Chain-of-Thought (CoT) reasoning into Vision-Language-Action (VLA) models. By utilizing a two-stage reasoning process—first predicting a subgoal image, then generating an action sequence—combined with hybrid attention and action chunking strategies, it achieves an 81.13% average success rate on the LIBERO benchmark, significantly outperforming existing methods.
Magma: A Foundation Model for Multimodal AI Agents: Magma unifies UI screenshots, robot data, and human manipulation videos into a single pre-training framework by labeling interactive regions on images (Set-of-Mark) and tracking motion trajectories in videos (Trace-of-Mark). This enables a single model to possess both multimodal understanding and cross-domain action prediction capabilities, achieving SOTA performance in both UI navigation and robotic manipulation.
ShowUI: One Vision-Language-Action Model for GUI Visual Agent: Based on Qwen2-VL-2B, ShowUI reduces redundant tokens by 33% and achieves a 1.4x speedup through UI-connected-graph-guided visual token selection. Combined with interleaved vision-language-action streaming and a curated 256K training dataset, it achieves state-of-the-art (SOTA) zero-shot accuracy of 75.1% on ScreenSpot with only 2B parameters.

Browse all 44 Robotics & Embodied AI papers →

🎮 Reinforcement Learning (4)¶

CALF: Communication-Aware Learning Framework for Distributed Reinforcement Learning: This paper proposes the CALF framework, which injects configurable network delay, jitter, and packet loss models into RL training. This reduces policy performance degradation by approximately 3-4 times when deployed on real distributed edge devices, revealing that network conditions represent an important but overlooked dimension in the sim-to-real gap.
Gazing at Rewards: Eye Movements as a Lens into Human and AI Decision-Making in Hybrid Visual Foraging: The Visual Forager (VF) model is proposed, which simulates human eye-movement strategies in hybrid visual search tasks through target feature modulation, target value modulation, and a ViT-based Actor-Critic decision-making network. It achieves a normalized score of 72.6% (compared to 87.4% for humans), with a saccade amplitude difference of only 0.01° (4.06° vs. 4.05° for humans), revealing for the first time how target value and prevalence jointly influence human search decisions.
GROVE: A Generalized Reward for Learning Open-Vocabulary Physical Skill: This paper proposes the GROVE framework, which constructs a generalized reward function by leveraging LLMs to generate physical constraints and VLMs to evaluate motion semantics in a complementary manner. By using a lightweight Pose2CLIP mapper to skip rendering and project poses directly into the semantic space, GROVE achieves open-vocabulary physical skill learning, yielding 8.4x faster training speed and a 22.2% improvement in motion naturalness compared to existing methods.
SkillMimic: Learning Basketball Interaction Skills from Demonstrations: SkillMimic is proposed, a purely data-driven framework that learns diverse basketball interaction skills from motion capture data using a unified HOI imitation reward (especially the innovative contact graph reward), and composes these skills using a high-level controller to complete complex long-horizon tasks such as continuous scoring.

🎁 Recommender Systems (1)¶

FineVQ: Fine-Grained User Generated Content Video Quality Assessment: This work constructs the first large-scale, fine-grained UGC video quality assessment database, FineVD (6,104 videos, 800k+ ratings, 6 dimensions), and proposes FineVQ, an LMM-based approach. FineVQ enables a single model to simultaneously perform quality rating, scoring, and attribution, achieving state-of-the-art performance on FineVD and other UGC-VQA datasets.

🔄 Self-Supervised Learning (30)¶

Order-Robust Class Incremental Learning: Graph-Driven Dynamic Similarity Grouping: GDDSG is proposed, which groups classes by similarity using graph coloring theory—ensuring classes within the same group are as dissimilar as possible (to reduce interference). Each group is learned independently using an NCM classifier and a LoRA adapter, achieving a 94.00% accuracy and only a 0.78% forgetting rate on CIFAR-100 10-step (prev. SOTA RanPAC: 90.50% / 3.49%).
OCRT: Boosting Foundation Models in the Open World with Object-Concept-Relation Triad: OCRT proposes a plug-and-play three-stage pipeline—Object (Slot Attention decoupling), Concept (importance filtering), and Relation (concept graph reasoning)—which significantly improves the accuracy of SAM on weakly-supervised medical/camouflaged segmentation and the robustness of CLIP under adversarial attacks, without altering the FM backbone.
Spectral State Space Model for Rotation-Invariant Visual Representation Learning: Proposed Spectral VMamba, which orders the patch traversal sequence using eigenvectors of the spectral graph Laplacian (instead of predefined scan lines) and combines it with a Rotation Feature Normalizer (RFN, aggregating features of 4 canonical rotations) to achieve 87.86% accuracy on miniImageNet with complete invariance to canonical rotations.
Transformers without Normalization: Discovered that the input-output mapping of LayerNorm exhibits a tanh-like shape, and proposed Dynamic Tanh (DyT) as a plug-and-play alternative to normalization layers: $\text{DyT}(x) = \gamma \odot \tanh(\alpha x) + \beta$. DyT achieves comparable or superior performance to LN across multiple tasks such as vision, language, diffusion, and speech.
From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling: Proposes a prototype-driven curriculum learning for MAE, which identifies "prototype" samples (representative images close to cluster centroids) in the dataset using K-means clustering. By using a temperature-controlled sampling strategy, the training smoothly transitions from prototypes to the full distribution, achieving an up to $8\times$ training acceleration (a 200-epoch prototype curriculum performs comparably to an 800-epoch standard MAE).
MAP: Unleashing Hybrid Mamba-Transformer Vision Backbone's Potential with Masked Autoregressive Pretraining: This work proposes Masked Autoregressive Pretraining (MAP). By utilizing a hierarchical pretraining objective that combines local MAE modeling with row-level autoregressive decoding, this work successfully pretrains hybrid Mamba-Transformer vision backbones for the first time, significantly outperforming individual MAE and AR strategies.
MetaWriter: Personalized Handwritten Text Recognition Using Meta-Learned Prompt Tuning: MetaWriter formulates the personalized adaptation of handwritten text recognition as a prompt tuning problem. By combining it with a Masked Autoencoder (MAE) self-supervised auxiliary task, it achieves label-free test-time adaptation. It optimizes prompt initialization via meta-learning to align the self-supervised loss with the recognition loss, achieving SOTA on IAM and RIMES by updating less than 1% of the parameters.
BoSS: A Best-of-Strategies Selector as an Oracle for Deep Active Learning: Proposes BoSS—a scalable active learning oracle strategy that generates candidate batches by ensembling multiple selection strategies, evaluates performance gain by freezing the backbone and retraining only the last layer, and selects the optimal batch. It showcases oracle performance on large-scale datasets like ImageNet for the first time, revealing that SOTA active learning strategies still have significant room for improvement.
Breaking the Tuning Barrier: Zero-Hyperparameters Yield Multi-Corner Analysis Via Learned Priors: Legacy hand-tuned priors are replaced with a pretrained Foundation Model (TabPFN) to achieve zero-hyperparameter tuning for circuit yield multi-corner analysis. By freezing the backbone to perform in-context learning, automatically transferring knowledge across corners, and integrating automatic feature selection (1152D to 48D), this method achieves SOTA accuracy (MRE down to 0.11%) on SRAM benchmarks while reducing verification costs by over 10x.
Do Your Best and Get Enough Rest for Continual Learning: Inspired by Ebbinghaus's forgetting curve theory, this paper proposes the View-Batch Model (VBM). By replacing multiple distinct samples in a batch with multiple augmented views (replay) of the same sample, VBM extends the recall interval by a factor of $V$ to an optimal range. Concurrently, it employs a one-to-many KL-divergence self-supervised loss to extract more knowledge from a single sample ("do your best"). Serving as a drop-in replacement, VBM consistently improves performance across various continual learning methods.

Browse all 30 Self-Supervised Learning papers →

📐 Optimization & Theory (12)¶

Federated Learning with Domain Shift Eraser: This paper proposes the FDSE method, which decomposes each network layer into a domain-free feature extractor (DFE, globally aggregated to enhance consensus) and a domain-specific shift eraser (DSE, personalized aggregated to retain local characteristics). Combined with BN consistency regularization, it achieves 76.77% on DomainNet (outperforming Ditto by 1.6%) and 91.58% on Office-Caltech10 (outperforming FedBN by 4.6%).
Conformal Prediction for Zero-Shot Models: Applying conformal prediction to zero-shot models to provide theoretically guaranteed uncertainty quantification and calibrated prediction sets for models like CLIP.
Model Poisoning Attacks to Federated Learning via Multi-Round Consistency: This work identifies that existing model poisoning attacks in federated learning cancel each other out due to cross-round directional inconsistency. It proposes PoisonedFL, which achieves a multi-round consistent attack through a fixed random direction vector, dynamic magnitude adjustment, and a hypothesis testing mechanism, bypassing 8 SOTA defenses without requiring any real client information.
Stop Walking in Circles! Bailing Out Early in Projected Gradient Descent: It is discovered that the PGD attack exhibits cyclic behavior on the $L_\infty$ ball for robust samples. Detecting cycles via hashing (PGD_CD) enables early stopping, which achieves an iteration reduction of up to 96% while maintaining identical robustness evaluation results.
Automatic Joint Structured Pruning and Quantization for Efficient Neural Network Training and Compression: Proposed the GETA framework to achieve automatic joint structured pruning and quantization-aware training: Quantization-Aware Dependency Graph (QADG) constructs a generic pruning search space + partially projected SGD guarantees layer-wise bit-width constraints + an interpretable joint learning strategy, achieving competitive or state-of-the-art compression performance on both CNNs and Transformers.
Convex Relaxation for Robust Vanishing Point Estimation in Manhattan World: GlobustVP introduces convex relaxation techniques to the Manhattan World vanishing point estimation problem for the first time. By formulating the joint estimation of vanishing point locations and line-to-VP associations as a QCQP and relaxing it into an SDP, it achieves a globally optimal and highly efficient solver (~50ms/image) robust to up to 70% outliers.
How to Merge Your Multimodal Models Over Time?: This paper proposes the TIME (Temporal Integration of Model Expertise) framework to systematically study the progressive merging of multimodal expert models over time. By defining a search space across three axes—initialization strategy, deployment strategy, and merging technique, the work uncovers key design principles for temporal model merging on the FoMo-in-Flux benchmark.
SCOPE: Semantic Coreset with Orthogonal Projection Embeddings for Federated Learning: SCOPE proposes a semantic coreset selection framework for federated learning. By leveraging zero-shot VLM (MobileCLIP-S2) to extract three scalar metrics (representation score, diversity score, and margin proximity), the server aggregates a global consensus to guide a two-stage pruning process (anomaly filtering + redundancy elimination) on clients. This achieves a 128-512× uplink bandwidth reduction and 7.72× speedup while maintaining competitive accuracy.
Test-Time Augmentation Improves Efficiency in Conformal Prediction: It is discovered that test-time data augmentation (TTA) can systematically improve the efficiency of conformal prediction. By learning augmentation weights on a calibration set to optimize the augmentation aggregation strategy, the prediction set size is reduced by 10-17% on ImageNet with ResNet-50 while strictly preserving the coverage guarantee.
Towards Stable and Storage-efficient Dataset Distillation: Matching Convexified Trajectory: This paper proposes the MCT (Matching Convexified Trajectory) method. By replacing SGD expert trajectories with a linear convex combination trajectory from random initialization to the optimal point, MCT simultaneously addresses the three major challenges of the traditional MTT method: trajectory instability, slow convergence, and high storage consumption.

Browse all 12 Optimization & Theory papers →

🔗 Causal Inference (4)¶

Adventurer: Optimizing Vision Mamba Architecture Designs for Efficiency: The Adventurer series of vision models is proposed, which adapts image inputs to a unidirectional causal scanning framework through two simple designs: "Heading Average token" and "Inter-layer Flipping". This allows the Mamba architecture to achieve 4-6x the training speed of existing Vision Mamba models on vision tasks, while maintaining comparable or even superior accuracy to ViT.
FG-VCE: Towards Fine-Grained Interpretability — Counterfactual Explanations for Misclassification with Saliency Partition: This paper proposes the FG-VCE (Fine-Grained Visual Contrastive Explanation) framework. By calculating feature point contributions via Shapley values, isolating local features using a saliency partition module, and employing an iterative counterfactual generation strategy, it achieves fine-grained counterfactual explanations at both the object and part levels for the first time. It reveals the specific causes of model misclassification: "which fine-grained features led to the error" and "which local regions dominated the prediction change."
Image Quality Assessment: Investigating Causal Perceptual Effects with Abductive Counterfactual Inference: This paper formulates Full-Reference Image Quality Assessment (FR-IQA) as a counterfactual inference problem. By using a Structural Causal Model (SCM), it distinguishes between the causal components related to perceptual quality and the noise components in deep features. This achieves training-free, backbone-agnostic robust quality prediction, obtaining competitive performance on multiple benchmark datasets.
Joint Scheduling of Causal Prompts and Tasks for Multi-Task Learning: Proposed the JSCPT (Joint Scheduling of Causal Prompts and Tasks) framework, which first designs Multi-Task Vision-Language Prompts (MTVLP) and eliminates spurious correlation features in prompts through causal intervention, and then adjusts learning order and weights using an adaptive task scheduler based on the dynamic changes in task relationships during training, achieving significant improvements across multiple multi-task visual recognition benchmarks.

🔬 Interpretability (22)¶

Scaling Vision Pre-Training to 4K Resolution: This paper proposes PS3 (Pre-training with Scale-Selective Scaling), which scales CLIP-style vision pre-training to 4K resolution with near-constant computational overhead by replacing global image contrastive learning with localized region and local caption contrastive learning. Combined with top-down/bottom-up patch selection mechanisms, the VILA-HD multimodal large language model is constructed, significantly outperforming GPT-4o and Qwen2.5-VL on high-resolution perception tasks.
Towards Human-Understandable Multi-Dimensional Concept Discovery: Proposed the HU-MCD framework, which replaces traditional segmentation methods with SAM to discover human-understandable visual concepts, coupled with a CNN-specific input masking scheme to reduce noise interference, achieving concept-level model explanations that balance understandability and faithfulness under the completeness framework of MCD.
Open Ad-Hoc Categorization with Contextualized Feature Learning: This paper proposes OAK (Open Ad-hoc Categorization with Contextualized Feature Learning). By introducing a few learnable context tokens into the input layer of a frozen CLIP model and combining CLIP's vision-language alignment objective with the visual clustering objective of GCD, the method achieves adaptive ad-hoc category discovery and context switching under few-labeled samples. It achieves an accuracy of 87.4% on novel classes of the Stanford Mood dataset, outperforming CLIP and GCD by over 50%.
Prompt-CAM: Making Vision Transformers Interpretable for Fine-Grained Analysis: Prompt-CAM is proposed to realize almost "free" interpretable fine-grained analysis. By injecting class-specific learnable prompt tokens into a pre-trained ViT, it utilizes the multi-head attention maps of the last layer to identify and localize critical traits that distinguish fine-grained categories.
Language Guided Concept Bottleneck Models for Interpretable Continual Learning: This paper introduces language-guided Concept Bottleneck Models (CBMs) into continual learning. It uses ChatGPT to generate human-interpretable concepts and the CLIP text encoder to encode concept embeddings, constructing a concept bottleneck layer. This provides transparent decision explanations while mitigating catastrophic forgetting, outperforming the SOTA on ImageNet-subset by 3.06%.
Attribute-formed Class-specific Concept Space: Endowing Language Bottleneck Model with Better Interpretability and Scalability: This paper proposes ALBM (Attribute-formed Language Bottleneck Model), which avoids spurious correlation reasoning by constructing an attribute-guided class-specific concept space, extracts fine-grained attribute features using visual attribute prompt learning, and automatically generates high-quality concept sets through a Description-Summary-Supplement (DSS) strategy, achieving better interpretability and scalability across 9 benchmarks.
Attribute-formed Class-specific Concept Space: Endowing Language Bottleneck Model with Better Interpretability and Scalability: This paper proposes the ALBM model, which replaces the class-shared concept space of existing Language Bottleneck Models (LBMs) with an Attribute-formed Class-specific Concept Space (ACCS) to address the issue of spurious cue reasoning and support cross-class generalization. Combined with Visual Attribute Prompt Learning (VAPL) to extract fine-grained attribute features, ALBM comprehensively outperforms existing interpretable classification methods on 9 few-shot benchmarks.
Differentiable Inverse Rendering with Interpretable Basis BRDFs: Proposes a differentiable inverse rendering method based on interpretable basis BRDFs, decomposing materials into combinations of physically meaningful basis functions to achieve interpretable material estimation.
Geometry-Guided Camera Motion Understanding in VideoLLMs: Proposes a complete framework spanning benchmark construction, diagnosis, and injection. By extracting camera motion cues from a 3D foundation model (VGGT) and injecting them into the VideoLLM via structured prompting, training-free camera motion perception enhancement is achieved.
Interpretable Image Classification via Non-parametric Part Prototype Learning: This paper proposes an interpretable image classification framework based on non-parametric prototype learning. It discovers semantically distinct object part prototypes by performing optimal transport clustering on self-supervised ViT features, addressing prototype redundancy issues in existing ProtoPNet methods, while introducing two new metrics, Distinctiveness and Comprehensiveness, to quantify explanation quality.

Browse all 22 Interpretability papers →

📦 Model Compression (71)¶

Towards Practical Real-Time Neural Video Compression: This paper proposes DCVC-RT, the first neural video codec to achieve 1080p real-time encoding and decoding on consumer-grade hardware with compression efficiency surpassing H.266/VTM. The core finding is that operational complexity (rather than computational complexity) acts as the actual speed bottleneck. Based on this, implicit temporal modeling and a single-scale low-resolution latent representation are designed, achieving encoding/decoding speeds of 125/113 fps on an A100 GPU while saving 21% bitrate.
Dataset Distillation with Neural Characteristic Function: A Minmax Perspective: The NCFM method is proposed, which reformulates dataset distillation as a minmax adversarial optimization problem by using the Neural Characteristic Function Difference (NCFD) parameterized by a neural network on the complex plane as a distribution distance metric. By aligning both phase (authenticity) and magnitude (diversity) information, it improves performance by up to 20.5% on ImageNet subsets while reducing GPU memory usage by over 300 times.
CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning: This paper proposes CL-LoRA, which designs a dual-adapter architecture (task-shared + task-specific LoRA). By combining knowledge distillation, gradient reassignment, and learnable block-wise weights, CL-LoRA achieves SOTA continual learning performance with only 0.3% trainable parameters.
JamMa: Ultra-lightweight Local Feature Matching with Joint Mamba: JamMa proposes an ultra-lightweight semi-dense feature matcher based on Joint Mamba. Through the JEGO scan-and-merge strategy, it achieves cross-view joint scanning, efficient four-way scanning, global receptive fields, and omnidirectional feature representation, achieving a superior performance-efficiency tradeoff compared to Transformer-based matchers with less than 50% of the parameters and FLOPs.
ECVC: Exploiting Non-Local Correlations in Multiple Frames for Contextual Video Compression: This paper proposes the ECVC video compression model, which captures non-local correlations among multiple reference frames through Multi-Frame Non-Local Contextual mining (MNLC) and Multi-Head Linear Cross-Attention (MHLCA). Combined with a Partial Cascaded Fine-Tuning Strategy (PCFS) to address the mismatch between training and testing sequence lengths, ECVC saves 10.5% and 11.5% bitrate under IP=32 and IP=-1 configurations compared to DCVC-FM, respectively.
LoRA Subtraction for Drift-Resistant Space in Exemplar-Free Continual Learning: LoRA-DRS proposes a "LoRA subtraction" operation—subtracting the LoRA weights of old tasks from the pre-trained weights before learning a new task to construct a Drift-Resistant Space (DRS). Subsequently, the LoRA of the new task is trained via gradient projection within this space, combined with an augmented triplet loss to enhance plasticity. This approach achieves SOTA performance in exemplar-free continual learning, with a particularly significant advantage in long task sequences.
Enhancing Dataset Distillation via Non-Critical Region Refinement: This paper proposes a three-stage framework, NRR-DD: using CAM to select low-confidence patches to initialize synthetic images, freezing critical regions while optimizing only non-critical regions to improve information density, and replacing 1000-dimensional soft labels with 2 distance values to achieve 500x storage compression. It achieves 46.1% accuracy on ImageNet-1K with IPC=10 (outperforming RDED by 25.7%), reducing soft label storage from 120GB to 0.2GB.
MuTri: Multi-view Tri-alignment for OCT to OCTA 3D Image Translation: This paper proposes MuTri, which introduces vector quantization (VQ) to the 3D volumetric translation task from OCT to OCTA for the first time. Through a two-stage training process—first pre-training VQ-VAEs for OCT and OCTA reconstruction to provide multi-view priors, and then utilizing multi-view guidance, namely contrastive semantic alignment (3D OCT/OCTA views) and vessel structure alignment (2D OCTA projection view), to guide the codebook learning of the translation VQ-VAE—the proposed method comprehensively outperforms state-of-the-art (SOTA) methods across three datasets.
Sampling Innovation-Based Adaptive Compressive Sensing: The SIB-ACS framework is proposed, which guides multi-stage adaptive sampling allocation through a "sampling innovation" criterion (measuring the reduction in reconstruction error brought by sampling increments) and designs a Principal Component Compressed Domain Network (PCCD-Net) for high-fidelity image reconstruction, significantly surpassing SOTA compressive sensing methods.
MobileMamba: Lightweight Multi-Receptive Visual Mamba Network: This work proposes MobileMamba, a lightweight visual network. Through a three-stage coarse-grained architectural design and the fine-grained MRFFI module (integrating Mamba global modeling, multi-kernel convolutional multi-scale perception, and Identity redundancy elimination), MobileMamba achieves an optimal balance between speed and accuracy on both classification and downstream high-resolution tasks.

Browse all 71 Model Compression papers →

🕸️ Graph Learning (7)¶

Coeff-Tuning: A Graph Filter Subspace View for Tuning Attention-Based Large Models: This paper reinterprets multi-head attention as a graph convolutional filter subspace, and linearly combines pre-trained attention maps by learning an extremely small set of subspace combination coefficients ($H \times H$ matrices). This breaks the convex hull constraint caused by the softmax function to expand the feature space, improving the performance of various PEFT methods in a plug-and-play manner at near-zero parameter cost.
DVHGNN: Multi-Scale Dilated Vision HGNN for Efficient Vision Recognition: This paper proposes DVHGNN, a vision backbone network that utilizes multi-scale dilated hypergraphs to capture high-order correlations among image patches. By employing clustering and Dilated Hypergraph Construction (DHGC) to extract multi-scale hyperedges, alongside dynamic hypergraph convolution for adaptive feature exchange, DVHGNN achieves an 83.1% top-1 accuracy on ImageNet-1K with 30.2M parameters, outperforming ViG-S by 1.0% and ViHGNN-S by 0.6%.
Hypergraph Vision Transformers: Images are More than Nodes, More than Edges: Proposed HgVT, which embeds a hierarchical bipartite hypergraph structure into ViTs. By processing primary image patch vertices and virtual vertices separately, constructing dynamic cosine adjacency, and utilizing a three-layer attention mechanism based on a hyperedge communication pool, HgVT captures high-order semantic relations among patches without clustering. On ImageNet-1K, HgVT-Ti achieves 76.2% accuracy with 7.7M parameters (outperforming ViHGNN-Ti by 1.9%) and reaches 73.23% mAP@10 in image retrieval.
Knowledge Bridger: Towards Training-Free Missing Modality Completion: This paper proposes Knowledge Bridger, a training-free framework for missing modality completion. By leveraging Large Multimodal Models (LMMs) to automatically mine multimodal knowledge and construct a knowledge graph, it guides the generation and ranking of missing modalities, surpassing existing methods in both general and medical OOD scenarios.
NN-Former: Rethinking Graph Structure in Neural Architecture Representation: NN-Former proposes a hybrid GNN-Transformer architecture predictor, revealing that existing methods overlook the topological information of "sibling nodes" (nodes sharing parent/child nodes). By introducing Adjacency-Sibling Multihead Attention (ASMA) and Bidirectional Graph Isomorphism FFN (BGIFFN), it achieves a Kendall's Tau of $0.877$/$0.890$ on NAS-Bench-101/201 and reduces the MAPE of latency prediction by 48-64%.
Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing: This paper proposes the VISA framework to debias video scene graph generation from both visual (Memory-Guided Sequence Modeling (MGSM) to reduce feature variance) and semantic (Iterative Relation Generator (IRG) to introduce hierarchical context and reduce dependence on biased priors) perspectives, significantly improving performance on tail categories on datasets like Action Genome.
Universal Scene Graph Generation: This paper proposes the Universal Scene Graph (USG) representation and its parser USG-Par, which generates a unified scene graph from arbitrary combinations of modalities (images, text, video, 3D) using a cross-modal object associator and text-centric scene contrastive learning, capturing both modality-invariant and modality-specific scene semantics.

📈 Time Series (5)¶

DejaVid: Encoder-Agnostic Learned Temporal Matching for Video Classification: This paper proposes DejaVid, an encoder-agnostic, lightweight approach for enhancing video classification. Instead of representing a video with a single embedding, DejaVid represents it as a variable-length Temporal Sequence of Embeddings (TSE). By learning importance weights for each time step and feature dimension, combined with an improved differentiable DTW algorithm for temporal alignment classification, it achieves SOTA results of 77.2% on SSV2 and 89.1% on K400 with an increase of only <1.8% parameters.
Competition-Aware CPC Forecasting with Near-Market Coverage: This work reformulates paid search CPC forecasting as a "forecasting under partially observable competition" problem. It approximates unobservable competitive states using three types of competition proxies: semantic neighborhoods (via Transformer embeddings), behavioral neighborhoods (via DTW alignment), and geographic intent. Evaluations on Google Ads data covering 1,811 keywords over 127 weeks demonstrate that competition-aware enhancements significantly outperform univariate and weak-context baselines in medium- to long-term forecasting (6/12 weeks).
L2GTX: From Local to Global Time Series Explanations: L2GTX proposes a completely model-agnostic global explanation method for time series classification. By aggregating Parameterized Event Primitives (PEPs) generated by LOMATCE, it constructs class-level global explanations, maintaining stable global fidelity ($R^2$) across six benchmark datasets.
Learning Extremely High Density Crowds as Active Matters: This paper models extremely high-density crowds ($\ge 5 \text{ people/m}^2$) as active matter, proposing a neural stochastic differential equation system that combines a novel "crowd material" stress model with Toner-Tu active forces. The system learns and predicts crowd dynamics directly from in-the-wild video optical flow using a hybrid Eulerian-Lagrangian CrowdMPM framework.
FLAVC: Learned Video Compression with Feature Level Attention: This work proposes FLAVC, which introduces a Feature-level Attention (FLA) module into the learned video compression (LVC) framework. By converting high-level local patch embeddings into one-dimensional batch-wise vectors and replacing traditional attention weights with a global context matrix, FLA achieves full-frame-level global perception. Combined with a Dense Overlapping Patcher and a hybrid Transformer-CNN encoder, FLAVC achieves state-of-the-art rate-distortion performance across four video compression datasets.

🏥 Medical Imaging (82)¶

Enhanced Contrastive Learning with Multi-view Longitudinal Data for Chest X-ray Report Generation: This paper proposes MLRG, a two-stage framework that integrates the spatial information of current multi-view images and the temporal information of historical longitudinal data during vision-language pre-training via multi-view longitudinal contrastive learning. It flexibly handles missing patient prior knowledge using tokenized absence encoding, achieving a 2.3% improvement in BLEU-4 on MIMIC-CXR and a 5.5% improvement in F1 on MIMIC-ABN.
Show and Segment: Universal Medical Image Segmentation via In-Context Learning: Iris is proposed as a framework that extracts task embeddings from reference image-label pairs using a lightweight task encoding module to guide target image segmentation. It adapts to new tasks without fine-tuning, achieving or exceeding the performance of task-specific models across 12 datasets, while demonstrating excellent generalization capabilities on 7 unseen datasets.
Federated Modality-specific Encoders and Partially Personalized Fusion Decoder for Multimodal Brain Tumor Segmentation: This work proposes FedMEPD, a federated learning framework that simultaneously addresses inter-modality heterogeneity and client personalization in multimodal MRI brain tumor segmentation through modality-specific encoders (globally federated) and a partially personalized fusion decoder. It achieves an average client mDSC of 75.70%/75.90% on BraTS 2018/2020.
SACB-Net: Spatial-Awareness Convolutions for Medical Image Registration: This paper proposes a 3D Spatial-Awareness Convolutional Block (SACB) that performs unsupervised clustering on feature maps and generates adaptive convolution kernels for different spatial regions. Combined with a pyramid flow estimator to achieve multi-scale deformation field composition, this method outperforms existing state-of-the-art (SOTA) methods on brain and abdomen CT registration tasks.
VISTA3D: A Unified Segmentation Foundation Model For 3D Medical Imaging: This paper proposes VISTA3D, the first unified 3D medical image segmentation foundation model. It simultaneously supports automatic segmentation of 127 classes, 3D interactive editing, and zero-shot segmentation. By utilizing a 3D supervoxel technology distilled from SAM, VISTA3D achieves state-of-the-art (SOTA) zero-shot performance and matches or exceeds specialized expert models on 14 datasets.
WISE: A Framework for Gigapixel Whole-Slide-Image Lossless Compression: To address the failure of existing lossless compression methods caused by the "information irregularity" (widespread high-frequency signals and high volatility) of WSI images, this paper proposes the WISE three-step compression framework (hierarchical projection coding $\rightarrow$ bitmap encoding $\rightarrow$ dictionary coding), achieving an average of 36x and up to 136x lossless compression.
DiN: Diffusion Model for Robust Medical VQA with Semantic Noisy Labels: This paper proposes the DiN framework, applying diffusion models to the noisy-label medical VQA (NM-VQA) scenario for the first time. Through a diffusion-based answer classifier, it screens answers from coarse to fine from a generative perspective. Combined with a noisy label refinement module to dynamically correct labels, DiN achieves an accuracy of 74.24% on VQA-RAD under 10% semantic noise, outperforming SNLC's 69.65%.
Automated Detection of Malignant Lesions in the Ovary Using Deep Learning Models and XAI: Detects ovarian cancer and its subtypes on histopathology images using 15 CNN variants (LeNet, ResNet, VGG, Inception), selects InceptionV3 (ReLU) as the optimal model (average 94.58% accuracy), and interprets model predictions using three XAI methods: LIME, SHAP, and Integrated Gradients.
Developing Foundation Models for Universal Segmentation from 3D Whole-Body Positron Emission Tomography: Constructed the largest PET segmentation dataset PETWB-Seg11K (11,041 whole-body PET cases + 59,831 segmentation masks) and proposed SegAnyPET—a foundation model for universal PET segmentation based on a 3D architecture + prompt engineering. It demonstrates strong zero-shot generalization capabilities across multi-center, multi-tracer, and multi-disease scenarios.
Diffusion-Based Feature Denoising and Using NNMF for Robust Brain Tumor Classification: Presents a brain tumor classification framework combining Non-Negative Matrix Factorization (NNMF) feature extraction, statistical feature selection, lightweight CNN classification, and diffusion-based feature space denoising. While maintaining ~85% clean accuracy, it improves robust accuracy under AutoAttack from 0.47% to 59.5%.

Browse all 82 Medical Imaging papers →

🧬 Computational Biology (7)¶

DiffVsgg: Diffusion-Driven Online Video Scene Graph Generation: DiffVsgg is proposed to model Video Scene Graph Generation (VSGG) as an iterative denoising problem along the temporal axis. It unifies object classification, box regression, and relation prediction using a shared feature embedding. Through latent diffusion models for spatial reasoning and using prior-frame predictions as conditioning for temporal reasoning, it achieves the first online VSGG and accomplishes comprehensive SOTA performance across all three evaluation protocols on Action Genome, surpassing DSG-DETR by 3.3 points in R@10.
Unsupervised Foundation Model-Agnostic Slide-Level Representation Learning: This work proposes Cobra, an unsupervised foundation model-agnostic (FM-agnostic) whole slide image (WSI)-level representation learning framework. It leverages embeddings from multiple pre-trained patch-level foundation models as feature-space augmentations, training a slide-level encoder using a Mamba-2 encoder and contrastive learning. Pre-trained on only 3,048 WSIs, Cobra outperforms existing slide encoders by at least +4.4% in average AUC across 15 downstream tasks.
Semantic and Expressive Variation in Image Captions Across Languages: This work systematically demonstrates significant distributional differences in semantic content (objects, relations, attributes) and expressive style (concreteness, tone, authenticity) in image captions across different languages. Multilingual caption sets provide richer visual information compared to monolingual ones (+46% objects, +66.1% relations, +66.8% attributes), providing empirical support for training vision models on multilingual data.
Multimodal Protein Language Models for Enzyme Kinetic Parameters: From Substrate Recognition to Conformational Adaptation: This paper proposes the ERBA adapter, which models enzyme kinetic prediction as a staged conditioning process of "substrate recognition $\rightarrow$ conformational adaptation". It injects substrate semantics via MRCA, fuses active-site 3D geometry via G-MoE, and preserves PLM priors via ESDA, consistently outperforming existing methods on three kinetic endpoints: kcat, Km, and Ki.
SHREC: A Spectral Embedding-Based Approach for Ab-Initio Reconstruction of Helical Molecules: This paper proposes the SHREC algorithm, which leverages the spectral embedding of the graph Laplacian to directly recover the projection angles of helical molecules from 2D cryo-EM projection images. Without requiring prior knowledge of helical symmetry parameters (rise/twist) and only requiring the axial point group symmetry $C_n$, SHREC achieves near-atomic resolution ab-initio helical reconstruction on multiple public datasets.
Synthetic Visual Genome: Proposes the SVG (Synthetic Visual Genome) data engine. Through a two-stage pipeline consisting of completing missing relationships on top of existing human annotations via GPT-4 (Stage 1) and Robin self-distillation + GPT-4 editing (Stage 2/SG-Edit), it generates a dense scene graph dataset with 146K images, 2.6M objects, and 5.6M relationships. The trained Robin-3B model outperforms same-sized models trained on over 300M instances using less than 3M instances, achieving a state-of-the-art (SOTA) score of 88.9 on referring expression comprehension.
Towards Spatio-Temporal World Scene Graph Generation from Monocular Videos: This paper proposes the World Scene Graph Generation (WSGG) task and the ActionGenome4D dataset, upgrading video scene graphs from frame-centric 2D representations to world-centric 4D representations. It requires models to perform 3D localization and relation prediction in the world coordinate system for all objects, including invisible ones that are occluded or out of view. Three complementary methods (PWG/MWAE/4DST) are proposed to explore different inductive biases for invisible object reasoning.

⚛️ Physics & Scientific Computing (8)¶

Accurate Differential Operators for Hybrid Neural Fields: This paper reveals that gradients and curvatures computed via automatic differentiation in hybrid neural fields (e.g., Instant NGP) suffer from severe high-frequency noise. It proposes a post-processing differential operator based on local polynomial fitting and a self-supervised fine-tuning method, reducing gradient and curvature errors by 4x, which significantly eliminates artifacts in rendering and physical simulations.
ATP: Adaptive Threshold Pruning for Efficient Data Encoding in Quantum Neural Networks: This paper proposes Adaptive Threshold Pruning (ATP) to adaptively prune low-information data features prior to quantum data encoding. By optimizing thresholds via L-BFGS-B, ATP achieves the highest accuracy in binary classification tasks across four datasets (MNIST, FashionMNIST, CIFAR, PneumoniaMNIST) while significantly reducing entanglement entropy.
DiffFNO: Diffusion Fourier Neural Operator: This paper proposes DiffFNO, which integrates the Weighted Fourier Neural Operator (WFNO) with a diffusion framework for arbitrary-scale super-resolution. It preserves critical high-frequency components through Mode Rebalancing, fuses frequency-domain and spatial-domain features using a Gated Fusion Mechanism, and accelerates inference with an adaptive-step ODE solver, outperforming existing methods by 2-4 dB in PSNR across multiple benchmarks.
Improve Representation for Imbalanced Regression through Geometric Constraints: This work is the first to study representation space uniformity in deep imbalanced regression (DIR). It proposes two geometric constraints, namely enveloping loss and homogeneity loss, to ensure that regression representations are uniformly distributed on the hypersphere. It also designs a surrogate-driven representation learning (SRL) framework to integrate global geometric constraints into mini-batch training, achieving SOTA on several DIR tasks such as age estimation.
KAC: Kolmogorov-Arnold Classifier for Continual Learning: First to apply Kolmogorov-Arnold Networks (KAN) to continual learning. By replacing B-splines with Radial Basis Functions (RBF) to construct the classifier KAC, consistent and significant performance gains are achieved across multiple continual learning methods (up to +20.70% on CUB200 40-step) with only 0.23M additional parameters.
Learning Phase Distortion with Selective State Space Models for Video Turbulence Mitigation: MambaTM is proposed as the first Mamba-based video atmospheric turbulence mitigation network. It reparameterizes the phase distortion traditionally represented by Zernike polynomials into Latent Phase Distortion (LPD) via a VAE, using LPD to guide the state transitions of SSMs. While maintaining linear complexity and a global receptive field, it achieves state-of-the-art restoration quality and nearly 2× inference speedup (55.4 FPS vs 32.7 FPS).
Towards Faithful Multimodal Concept Bottleneck Models: Proposes f-CBM, a faithful multimodal Concept Bottleneck Model framework based on CLIP. By jointly addressing concept detection accuracy and information leakage via a differentiable leakage loss and a Kolmogorov-Arnold Network prediction head, it achieves the optimal trade-off among task accuracy, concept detection, and leakage.
FreeGave: 3D Physics Learning from Dynamic Videos by Gaussian Velocity: Ours proposes FreeGave, a general framework for learning 3D scene geometry, appearance, and physical velocity from multi-view dynamic videos. By introducing a learnable physics code for each 3D Gaussian kernel and designing a divergence-free velocity field parameterization, FreeGave achieves accurate future frame extrapolation without relying on PINN losses or target priors.

🌍 Earth Science (1)¶

GeoChemAD: Benchmarking Unsupervised Geochemical Anomaly Detection for Mineral Exploration: This paper proposes the GeoChemAD open-source benchmark dataset (comprising 8 subsets covering multiple regions, sampling sources, and target elements) and the GeoChemFormer framework. By employing spatial context self-supervised pre-training and elemental dependency modeling, it achieves unsupervised geochemical anomaly detection and obtains state-of-the-art AUC across all subsets.

📡 Signal & Communications (5)¶

Breaking the Low-Rank Dilemma of Linear Attention: This paper theoretically reveals that the fundamental cause of linear attention's performance lagging behind Softmax attention is the low-rank bottleneck of output features. It proposes Rank-Augmented Linear Attention (RALA), which utilizes two complementary strategies—enhancing KV buffer rank and output feature rank—to match or even surpass the performance of Softmax attention while maintaining linear complexity.
DiTASK: Multi-Task Fine-Tuning with Diffeomorphic Transformations: Proposes DiTASK, which utilizes continuous piecewise-affine (CPAB) diffeomorphic transformations to smoothly transform the singular values of pretrained weight matrices while keeping the singular vectors unchanged. It achieves full-rank update multi-task fine-tuning with only about 32 parameters per layer, outperforming MTLoRA by 26.27% relative improvement with 75% fewer parameters on PASCAL MTL.
Neural Video Compression with Context Modulation: Proposed the DCMVC framework, which modulates temporal context in two steps: flow orientation and context compensation. By fully utilizing reference information in both the pixel domain and the feature domain, it achieves compression performance that saves an average of 22.7% bitrate compared to H.266/VVC and 10.1% bitrate compared to the previous SOTA, DCVC-FM.
ABC-Former: Auxiliary Bimodal Cross-domain Transformer with Interactive Channel Attention: This paper proposes ABC-Former, which introduces CIELab color space and RGB histograms as auxiliary bimodal information. It utilizes a cross-domain Transformer and an Interactive Channel Attention (ICA) module to achieve cross-modal transfer of global color knowledge, achieving SOTA performance in sRGB white balance correction tasks. It is also extended to ABC-FormerM to handle mixed illumination scenarios.
Continuous Space-Time Video Resampling with Invertible Motion Steganography: An Invertible Motion Steganography Module (IMSM) is proposed to embed motion information into low-frame-rate frames during video temporal downsampling, and accurately restore motion details via inverse transformation during upsampling. It supports continuous (non-integer) space-time resampling factors, significantly improving reconstruction quality while preserving the visual quality of downsampled frames.

👥 Social Computing (6)¶

As Language Models Scale, Low-order Linear Depth Dynamics Emerge: By treating the depth dimension of Transformers as a discrete-time dynamical system, this paper finds that a linear state-space surrogate model of just 32 dimensions can predict inter-layer sensitivity curves with high precision (Spearman up to 0.99) within a given context. Surprisingly, as the model scales, the low-order linear surrogate becomes even more accurate—unveiling a new scaling law.
Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification: This paper proposes Classifier-guided CLIP Distillation (CCD), which achieves unsupervised multi-label classification performance on par with fully supervised methods (90.1% mAP on VOC12) without any manual annotations by leveraging two core techniques: CAM-guided local view label aggregation and CLIP prediction debiasing.
Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers: Proposes C2B (Classifier-to-Bias), the first framework to automatically detect biases in pre-trained visual classifiers using only the textual descriptions of the classification task (without any labeled data). By leveraging LLMs to generate class-specific bias candidates, creating retrieval queries to collect image datasets, and finally calculating bias scores, C2B outperforms supervised SOTA bias detection methods on CelebA and ImageNet-X.
Learning from Neighbors: Category Extrapolation for Long-Tail Learning: It is discovered that finer-grained category division naturally mitigates the impact of long-tail imbalance. This paper proposes using LLMs to discover fine-grained auxiliary categories semantically related to existing ones, web crawlers to collect images, and a Neighbor-Silencing Loss to prevent auxiliary classes from dominating. This achieves a 16-percentage-point improvement ($41.4\% \to 57.4\%$) on Few-shot classes in ImageNet-LT.
Project-Probe-Aggregate: Efficient Fine-Tuning for Group Robustness: This paper proposes Project-Probe-Aggregate (PPA), a three-step framework that improves the group robustness of foundation models without group annotations, using less than 0.01% of trainable parameters. PPA projects features to remove class proxies and amplify bias, probes group labels corrected with group priors, and aggregates group weights.
Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples: Based on abstract: Deep learning models are known to often learn features that spuriously correlate with the class label during training but are irrelevant to the prediction task. Existing methods typically address this issue by annotating potential spurious attributes, or filtering spurious features based on some emp

📂 Others (66)¶

TraF-Align: Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception: Proposes the TraF-Align framework, which learns the spatiotemporal flow path of features by predicting object motion trajectories at the feature level. It generates temporally ordered sampling points along the trajectory to guide the current-timestamp query to relevant historical features, achieving precise feature alignment in asynchronous multi-agent perception. It achieves state-of-the-art (SOTA) performance on two real-world datasets: V2V4Real and DAIR-V2X-Seq.
Exploring Contextual Attribute Density in Referring Expression Counting (CAD-GD): The concept of Contextual Attribute Density (CAD) is proposed to enhance referring expression counting. By incorporating three modules—a U-shape density estimator, CAD attention, and dynamic query initialization—the approach reduces counting errors on the REC-8K dataset by approximately 30% compared to GroundingREC (MAE decreases from 6.80 to 5.43).
EDM: Equirectangular Projection-Oriented Dense Kernelized Feature Matching: EDM is proposed as the first learning-based dense feature matching method for Equirectangular Projection (ERP) panoramic images. It addresses the polar distortion of ERP through a Spherical Space Alignment Module (SSAM, utilizing spherical positional encoding with 3D Cartesian coordinates + Gaussian Process regression) and geodesic flow refinement. On Matterport3D, it outperforms DKM by 26.72% in AUC@5°, and on Stanford2D3D by 42.62%.
Do ImageNet-trained Models Learn Shortcuts? The Impact of Frequency Shortcuts on Generalization: This paper proposes a Hierarchical Frequency Shortcut Search (HFSS) method to efficiently discover frequency shortcuts learned by CNNs and Transformers at the ImageNet-1K scale for the first time (permitting correct classification with only 5% of frequencies). It reveals that frequency shortcuts are surprisingly beneficial in texture-preserving OOD tests but detrimental in stylized tests (IN-R/IN-S), pointing out that existing OOD evaluation frameworks overlook the impact of frequency shortcuts.
Scene-Agnostic Pose Regression for Visual Localization: Proposed a new task paradigm called "Scene-Agnostic Pose Regression" (SPR), which regresses the relative poses of subsequent frames using the first frame of the sequence as the coordinate origin. This avoids the need for retraining in APR, database retrieval in RPR, and cumulative drift in VO. A large-scale dataset, 360SPR, containing 200K panorama images, and a dual-branch SPR-Mamba model are established.
Foundations of the Theory of Performance-Based Ranking: This paper establishes a rigorous mathematical foundation for performance-based ranking based on probability theory and order theory. It proposes a general framework consisting of 6 pillars and 3 axioms, defines a parameterized family of "ranking scores," and demonstrates in binary classification tasks that metrics such as accuracy, TPR, TNR, PPV, and F-score satisfy these axioms, whereas commonly used metrics like MCC and geometric mean are unsuitable for ranking.
Instance-wise Supervision-level Optimization in Active Learning: This paper proposes the ISO (Instance-wise Supervision-level Optimization) framework. In active learning, it not only selects which samples to annotate but also automatically determines the optimal annotation level (exact vs. coarse labels) for each sample. Through a value-cost ratio (VCR) and a diversity-aware batch selection algorithm, it achieves over 10% higher accuracy than traditional active learning under a fixed budget constraint.
HotSpot: Signed Distance Function Optimization with an Asymptotically Sufficient Condition: This paper proposes HotSpot, which utilizes the classical relationship between the screened Poisson equation and distance fields to design a new heat loss. It provides an asymptotically sufficient condition for optimizing neural signed distance functions, ensuring that the implicit function converges to the true distance field. It significantly outperforms existing methods in 2D/3D surface reconstruction with complex topologies.
Image Reconstruction from Readout-Multiplexed Single-Photon Detector Arrays: This paper formalizes the multiphoton coincidence resolution problem in row-column readout-multiplexed single-photon detector arrays as an inverse imaging problem. It proposes a probabilistic Multiphoton Estimator (ME) capable of resolving the spatial locations of up to 4 concurrent incident photons. Compared to traditional methods, ME achieves a 3-4 dB PSNR improvement on a 32×32 array and reduces the required frame count by approximately 4 times.
Practical Solutions to the Relative Pose of Three Calibrated Cameras: This paper addresses the classic challenge of relative pose estimation for three calibrated cameras from four point correspondences in three views (4p3v). It proposes practical solutions based on approximate geometry—using affine camera approximations or mean point correspondence approximations to estimate the relative pose of the first two cameras, and then registering the third camera via P3P. Combined with local optimization, this approach achieves SOTA accuracy on real-world data.

Browse all 66 Others papers →