AAAI 2026 Papers

Video generation using Large Language Models (LLMs) has shown promising potential, effectively leveraging the extensive LLM infrastructure to provide a unified framework for multimodal understanding and content generation. However, these methods face critical challenges, i.e., token redundancy and inefficiencies arising from long sequences, which constrain their performance and efficiency compared to diffusion-based approaches. In this study, we investigate the impact of token redundancy in LLM-based video generation by information-theoretic analysis and propose Vision Representation Compression (VRC), a novel framework designed to achieve more in both performance and efficiency with less video token representations. VRC introduces learnable representation compressor and decompressor to compress video token representations, enabling autoregressive next-sequence prediction in a compact latent space. Our approach reduces redundancy, shortens token sequences, and improves model's ability to capture underlying video structures. Our experiments demonstrate that VRC reduces token sequence lengths by a factor of 4, achieving more than 9~14x acceleration in inference while maintaining performance comparable to state-of-the-art video generation models. VRC not only accelerates the inference but also significantly reduces memory requirements during both model training and inference.

Drag-Based Image Editing (DBIE), which allows users to manipulate images by directly dragging objects within them, has recently attracted much attention from the community. However, it faces two key challenges: (i) point-based drag is often highly ambiguous and difficult to align with user intentions; (ii) current DBIE methods primarily rely on alternating between motion supervision and point tracking, which is not only cumbersome but also fails to produce high-quality results. These limitations motivate us to explore DBIE from a new perspective---unifying it as a Latent Region Optimization (LRO) problem that aims to use region-level geometric transformations to optimize latent code to realize drag manipulation. Thus, by specifying the areas and types of geometric transformations, we can effectively address the ambiguity issue. We also propose a simple yet effective editing framework, dubbed DragNeXt. It solves LRO through Progressive Backward Self-Intervention (PBSI), simplifying the overall procedure of the alternating workflow while further enhancing quality by fully leveraging region-level structure information and progressive guidance from intermediate drag states. We validate DragNeXt on our NextBench, and extensive experiments demonstrate that our proposed method can significantly outperform existing approaches.

Despite the rapid progress of Vision Language Models (VLMs), existing benchmarks still concentrate on coarse-grained object recognition or simple relational reasoning, leaving the fine-grained and higher-order reasoning abilities of these systems largely unexamined. To bridge this critical evaluation gap, we introduce EmojiGrid, a novel diagnostic benchmark specifically designed to probe these fine-grained and higher-order skills. Leveraging the universal and semantically rich nature of emojis, we synthesize a grid‑based visual dataset paired with 29,000+ QA pairs. Each pair is explicitly anchored in a three-level cognitive taxonomy comprising (i) Perception and Information Extraction, (ii) Relational and Structural Reasoning, and (iii) Abstraction and Advanced Cognition. These dimensions further decompose into nine categories covering a broad range of cognitive skills, including counting, spatial relations, compositional logic, semantic sentiment, and related higher-order reasoning tasks. Our extensive evaluation of 25 state-of-the-art open-source and proprietary VLMs reveals a significant performance gap between foundational perceptual tasks and higher-level cognitive abilities, particularly in abstraction and advanced emotional reasoning. Notably, all models struggle with compositional logic, spatial consistency, and especially emotional and semantic understanding. EmojiGrid provides a quantifiable, fine-grained benchmark to diagnose VLM limitations and guides future progress toward models that can truly perceive, reason about, and interpret complex, symbol-rich visual scenes.

Rectification flow Transformers (RFTs) have shown promising performance in diffusion-based image synthesis but are typically confined to lower-resolution scenarios, limiting their ability to generate high-resolution images. Existing resolution extrapolation approaches often suffer from excessive computational overhead, resulting in prolonged inference times. We propose LookFlow, a training-free high-resolution synthesis framework that accelerates inference while preserving visual quality. Building on pretrained text-to-image RFTs, LookFlow employs a dynamic lookahead guidance flow mechanism to refine high-resolution velocity predictions by leveraging multi-timestep lookahead information extracted from a low-resolution flow. Additionally, reusing temporally similar features across consecutive timesteps drastically reduces computation and significantly decreases inference time overhead. Extensive experiments on COCO demonstrate that LookFlow robustly scales resolutions from 4× to 25×, achieving up to a maximum speedup of 2.01× while maintaining competitive visual fidelity.

Keyframe selection has become essential for video understanding with vision-language models (VLMs) due to limited input tokens and the temporal sparsity of relevant information across video frames. Video understanding often relies on effective keyframes that are not only informative but also causally decisive. To this end, we propose Reinforced Causal Search with Information Bottleneck (ReaSon), a framework that formulates keyframe selection as an optimization problem with the help of a novel Causal Information Bottleneck (CIB), which explicitly defines keyframes as those satisfying both predictive sufficiency and causal necessity. Specifically, ReaSon employs a learnable policy network to select keyframes from a visually relevant pool of candidate frames to capture predictive sufficiency, and then assesses causal necessity via counterfactual interventions. Finally, a composite reward aligned with the CIB principle is designed to guide the selection policy through reinforcement learning. Extensive experiments on NExT-QA, EgoSchema, and Video-MME demonstrate that ReaSon consistently outperforms existing state-of-the-art methods under limited-frame settings, validating its effectiveness and generalization ability.

We present OpenDriveVLA, a Vision-Language Action (VLA) model designed for end-to-end autonomous driving, built upon open-source large language models. OpenDriveVLA generates spatially-grounded driving actions by leveraging multimodal inputs, including both 2D and 3D instance-aware visual representations, ego vehicle states, and language commands. To bridge the modality gap between driving visual representations and language embeddings, we introduce a hierarchical vision-language alignment process, projecting both 2D and 3D structured visual tokens into a unified semantic space. Furthermore, we incorporate structured agent–environment–ego interaction modeling into the autoregressive decoding process, enabling the model to capture fine-grained spatial dependencies and behavior-aware dynamics critical for reliable trajectory planning. Extensive experiments on the nuScenes dataset demonstrate that OpenDriveVLA achieves state-of-the-art results across open-loop trajectory planning and driving-related question-answering tasks. Qualitative analyses further illustrate its superior capability to follow high-level driving commands and robustly generate trajectories under challenging scenarios, highlighting its potential for next-generation end-to-end autonomous driving.

3D single object tracking (SOT) in LiDAR point clouds is a critical task in computer vision and autonomous driving. Despite great success having been achieved, the inherent sparsity of point clouds introduces a dual-redundancy challenge that limits existing trackers: (1) vast spatial redundancy from background noise impairs accuracy, and (2) informational redundancy within the foreground hinders efficiency. To tackle these issues, we propose CompTrack, a novel end-to-end framework that systematically eliminates both forms of redundancy in point clouds. First, CompTrack incorporates a Spatial Foreground Predictor (SFP) module to filter out irrelevant background noise based on information entropy, addressing spatial redundancy. Subsequently, its core is an Information Bottleneck-guided Dynamic Token Compression (IB-DTC) module that eliminates the informational redundancy within the foreground. Theoretically grounded in low-rank approximation, this module leverages an online SVD analysis to adaptively compress the redundant foreground into a compact and highly informative set of proxy tokens. Extensive experiments on KITTI, nuScenes and Waymo datasets demonstrate that CompTrack achieves top-performing tracking performance with superior efficiency, running at a real-time 90 FPS on a single RTX 3090 GPU.

Class-agnostic 3D instance segmentation tackles the challenging task of segmenting all object instances, including previously unseen ones, without semantic class reliance. Current methods struggle with generalization due to the scarce annotated 3D scene data or noisy 2D segmentations. While synthetic data generation offers a promising solution, existing 3D scene synthesis methods fail to simultaneously satisfy geometry diversity, context complexity, and layout reasonability, each essential for this task. To address these needs, we propose an Adapted 3D Scene Synthesis pipeline for class-agnostic 3D Instance SegmenTation, termed as ASSIST-3D, to synthesize proper data for model generalization enhancement. Specifically, ASSIST-3D features three key innovations, including 1) Heterogeneous Object Selection from extensive 3D CAD asset collections, incorporating randomness in object sampling to maximize geometric and contextual diversity; 2) Scene Layout Generation through LLM-guided spatial reasoning combined with depth-first search for reasonable object placements; and 3) Realistic Point Cloud Construction via multi-view RGB-D image rendering and fusion from the synthetic scenes, closely mimicking real-world sensor data acquisition. Experiments on ScanNetV2, ScanNet++, and S3DIS benchmarks demonstrate that models trained with ASSIST-3D-generated data significantly outperform existing methods. Further comparisons underscore the superiority of our purpose-built pipeline over existing 3D scene synthesis approaches.

Accurate point tracking in surgical environments remains challenging due to complex visual conditions, including smoke occlusion, specular reflections, and tissue deformation. While existing surgical tracking datasets provide coordinate information, they lack the semantic context necessary to understand tracking failure mechanisms. We introduce VL-SurgPT, the first large-scale multimodal dataset that bridges visual tracking with textual descriptions of point status in surgical scenes. The dataset comprises 908 in vivo video clips, including 754 for tissue tracking (17,171 annotated points across five challenging scenarios) and 154 for instrument tracking (covering seven instrument types with detailed keypoint annotations). We establish comprehensive benchmarks using eight state-of-the-art tracking methods and propose TG-SurgPT, a text-guided tracking approach that leverages semantic descriptions to improve robustness in visually challenging conditions. Experimental results demonstrate that incorporating point status information significantly improves tracking accuracy and reliability, particularly in adverse visual scenarios where conventional vision-only methods struggle. By bridging visual and linguistic modalities, VL-SurgPT enables the development of context-aware tracking systems crucial for advancing computer-assisted surgery applications that can maintain performance even under challenging intraoperative conditions.

We challenge the assumption that complex instruction-guided segmentation tasks necessitate equally complex and explicit supervision. This paper introduces RISE (Reasoning via Implicit Self-supervised Emergence), a framework that learns intricate compositional reasoning, spanning spatial relations to world knowledge, without a single ground-truth mask. To achieve this, RISE employs reinforcement learning with GRPO guided by a single, strikingly simple reward: the semantic alignment score between the textual instruction and the predicted image region. Our primary discovery is the implicit emergence of a high-quality chain-of-thought process from this minimalist signal. Within a structured format, the model autonomously learns to understand instructions by accessing its latent knowledge, inferring spatial relationships—capabilities inherent in its architecture but unlocked by our simple objective. Remarkably, our emergent reasoning yields highly competitive results: RISE achieves 58.7 gIoU on the ReasonSeg benchmark, on par with methods using geometric rewards. Furthermore, we show extreme data efficiency: a variant trained on only 2,000 ImageNet-label pairs establishes a new state-of-the-art for annotation-free referring segmentation with 79.6 cIoU on RefCOCO.

The practical deployment of infrared imaging is hindered by its inherent output of low-resolution (LR) images. While the super-resolution (SR) technique is a promising remedy, we discover two major challenges concerning infrared image SR: preserving accurate thermal distributions, which are fundamental to infrared imaging, and addressing the ambiguity of high-frequency elements compared to visible images. To tackle these issues, we propose ThesIS, a tailored framework that utilizes Thermal-Physics guidance and dynamic high-frequency amplification for Infrared image Super-resolution to produce high-resolution (HR) images with accurate physical properties and delicate visual details. Specifically, Thermal Regularization is introduced to reconstruct the accurate thermal radiation distribution via the introduced Infrared Radiation Intensity Alignment Loss, mitigating the adverse effects of complex degradations while conducting initial upscaling. Additionally, we design a guidance mechanism to counter the randomness of the diffusion model, further refining the preservation of physical information. The proposed Dynamic High-Frequency Amplification effectively strengthens the ambiguous high-frequency information present in infrared images, leading to improved texture details and superior visual quality. Extensive experiments demonstrate that ThesIS successfully recovers accurate thermal information while delivering visually satisfying results with state-of-the-art performance. Furthermore, we introduce the InfraredSR dataset, which comprises 39,833 images at a resolution of 512 × 512, hoping to advance research in this field.

Person re-identification (ReID) aims to retrieve target pedestrian images given either visual queries (image-to-image, I2I) or textual descriptions (text-to-image, T2I). Although both tasks share a common retrieval objective, they pose distinct challenges: I2I emphasizes discriminative identity learning, while T2I requires accurate cross-modal semantic alignment. Existing methods often treat these tasks separately, which may lead to representation entanglement and suboptimal performance. To address this, we propose a unified framework named Hierarchical Prompt Learning (HPL), which leverages task-aware prompt modeling to jointly optimize both tasks. Specifically, we first introduce a Task-Routed Transformer, which incorporates dual classification tokens into a shared visual encoder to route features for I2I and T2I branches respectively. On top of this, we develop a hierarchical prompt generation scheme that integrates identity-level learnable tokens with instance-level pseudo-text tokens. These pseudo-tokens are derived from image or text features via modality-specific inversion networks, injecting fine-grained, instance-specific semantics into the prompts. Furthermore, we propose a Cross-Modal Prompt Regularization strategy to enforce semantic alignment in the prompt token space, ensuring that pseudo-prompts preserve source-modality characteristics while enhancing cross-modal transferability. Extensive experiments on multiple ReID benchmarks validate the effectiveness of our method, achieving state-of-the-art performance on both I2I and T2I tasks.

Entity hallucination poses a major challenge in radiology report generation (RRG), particularly for 3D CT scans where complex spatial contexts amplify factual errors. To address this, medical entity phrases serve as key carriers for multi-modal prompting, integrating expert knowledge into the vision-language model. Current methods use unified cross-attention for volume-phrase alignment, failing to account for anatomical specificity during the alignment process. In this work, we introduce the Dual-stream Entity Alignment Reporting network (DEAR) that separately models organ and lesion entities to resolve anatomical bias. Specifically, the dual-stream entity aligner is designed to partition medical entity phrases into organ and lesion streams, feeding them into separate cross-attention blocks in parallel to achieve fine-grained volume–phrase alignment. For structurally regular and spatially stable organ entities, an organ-guided cross-attention (OGCA) block is proposed to enforce structural consistency by retrieving the top-k voxel tokens via volume–phrase similarity and preserving spatial connectivity through morphological dilation. Meanwhile, a lesion-guided cross-attention (LGCA) block is introduced for structurally irregular and spatially variable lesion entities, enhancing anomaly sensitivity through phrase-weighted attention and refining discriminative boundaries via 3D residual Laplacian filtering. Experiments demonstrate that DEAR significantly reduces entity hallucinations and improves clinical factuality in 3D RRG benchmarks.

Understanding anomalous human behaviors at a fine-grained level remains a major challenge in complex scenarios. Existing video anomaly understanding (VAU) methods often rely on coarse frame-level cues or overlook structured modeling of individual actions, limiting their capacity for reasoning about human interactions and accountability. To address these challenges, we propose TargetVAU, a multimodal anomaly-aware reasoning framework designed for individual-level anomaly recognition and explanation. TargetVAU first extracts both global-level and human-centric visual features using a frozen Vision Transformer (ViT) encoder. An Anomaly-focused Temporal Sampler is then employed to select behaviorally informative frames via a density-aware strategy guided by predicted anomaly scores. A Spatio-Temporal Interaction Graph is constructed to explicitly model interactions among individuals across time and space. These structured representations are fused with prompt embeddings via a frozen Q-Former to form a unified semantic representation. Finally, a large language model fine-tuned with low-rank adaptation (LoRA) performs instruction-guided reasoning to identify anomalous individuals and generate natural language explanations. Extensive experiments on UCCD and HIVAU-70K demonstrate that TargetVAU significantly outperforms existing methods in both accuracy and interpretability, advancing the state of individual-level anomaly understanding in surveillance videos.

3D point cloud classification is a fundamental task in safety-critical applications such as autonomous driving, robotics, and augmented reality. However, recent studies reveal that point cloud classifiers are vulnerable to structured adversarial perturbations and geometric corruptions, posing risks to their deployment in safety-critical scenarios. Existing certified defenses limit point-wise perturbations but overlook subtle geometric distortions that preserve individual points yet alter the overall structure, potentially leading to misclassification. In this work, we propose FreqCert, a novel certification framework that departs from conventional spatial domain defenses by shifting robustness analysis to the frequency domain, enabling structured certification against global l2-bounded perturbations. FreqCert first transforms the input point cloud via the graph Fourier transform (GFT), then applies structured frequency-aware subsampling to generate multiple sub-point clouds. Each sub-cloud is independently classified by a standard model, and the final prediction is obtained through majority voting, where sub-clouds are constructed based on spectral similarity rather than spatial proximity, making the partitioning more stable under l2 perturbations and better aligned with the object’s intrinsic structure. We derive a closed-form lower bound on the certified l2 robustness radius and prove its tightness under minimal and interpretable assumptions, establishing a theoretical foundation for frequency domain certification. Extensive experiments on the ModelNet40 and ScanObjectNN datasets demonstrate that FreqCert consistently achieves higher certified accuracy and empirical accuracy under strong perturbations. Our results suggest that spectral representations provide an effective pathway toward certifiable robustness in 3D point cloud recognition.

Recovering point clouds involves the sequential process of sampling and restoration, yet existing methods struggle to effectively leverage both topological and geometric attributes. To address this, we propose an end-to-end architecture named TopGeoFormer, which maintains these critical properties throughout the sampling and restoration phases. First, we revisit traditional feature extraction techniques to yield topological embedding using a continuous mapping of relative relationships between neighboring points, and integrate it in both phases for preserving the structure of the original space. Second, we propose the InterTwining Attention to fully merge topological and geometric embeddings, which queries shape with local awareness in both phases to form a learnable 3D shape context facilitated with point-wise, point-shape-wise, and intra-shape features. Third, we introduce a full geometry loss and a topological constraint loss to optimize the embeddings in both Euclidean and topological spaces. The geometry loss uses inconsistent matching between coarse-to-fine generations and targets for reconstructing better geometric details, and the constraint loss limits embedding variances for better approximation of the topological space. In experiments, we comprehensively analyze the circumstances using the conventional and learning-based sampling/upsampling/recovery algorithms. The quantitative and qualitative results demonstrate that our method significantly outperforms existing sampling and recovery methods.

Estimating the 3D poses of hands and objects from a single RGB image is a fundamental yet challenging problem, with broad applications in augmented reality and human-computer interaction. Existing methods largely rely on visual cues alone, often producing results that violate physical constraints such as interpenetration or non-contact. Recent efforts to incorporate physics reasoning typically depend on post-optimization or non-differentiable physics engines, which compromise visual consistency and end-to-end trainability. To overcome these limitations, we propose a novel framework that jointly integrates visual and physical cues for hand-object pose estimation. This integration is achieved through two key ideas: 1) joint visual-physical cue learning: The model is trained to extract 2D visual cues and 3D physical cues, thereby enabling more comprehensive representation learning for hand-object interactions; 2) candidate pose aggregation: A novel refinement process that aggregates multiple diffusion-generated candidate poses by leveraging both visual and physical predictions, yielding a final estimate that is visually consistent and physically plausible. Extensive experiments demonstrate that our method significantly outperforms existing state-of-the-art approaches in both pose accuracy and physical plausibility.

The Dense Audio-Visual Event Localization (DAVEL) task aims to temporally localize events in untrimmed videos that occur simultaneously in both the audio and visual modalities. This paper explores DAVEL under a new and more challenging weakly-supervised setting (W-DAVEL task), where only video-level event labels are provided and the temporal boundaries of each event are unknown. We address W-DAVEL by exploiting cross-modal salient anchors, which are defined as reliable timestamps that are well predicted under weak supervision and exhibit highly consistent event semantics across audio and visual modalities. Specifically, we propose a Mutual Event Agreement Evaluation module, which generates an agreement score by measuring the discrepancy between the predicted audio and visual event classes. Then, the agreement score is utilized in a Cross-modal Salient Anchor Identification module, which identifies the audio and visual anchor features through global-video and local temporal window identification mechanisms. The anchor features after multimodal integration are fed into an Anchor-based Temporal Propagation module to enhance event semantic encoding in the original temporal audio and visual features, facilitating better temporal localization under weak supervision. We establish benchmarks for W-DAVEL on both the UnAV-100 and ActivityNet1.3 datasets. Extensive experiments demonstrate that our method achieves state-of-the-art performance.

Referring Audio-Visual Segmentation (Ref-AVS) aims to segment target objects in audible videos based on given reference expressions. Prior works typically rely on learning latent embeddings via multimodal fusion to prompt a tunable SAM/SAM2 decoder for segmentation, which requires strong pixel-level supervision and lacks interpretability. From a novel perspective of explicit reference understanding, we propose TGS-Agent, which decomposes the task into a Think-Ground-Segment process, mimicking the human reasoning procedure by first identifying the referred object through multimodal analysis, followed by coarse-grained grounding and precise segmentation. To this end, we first propose Ref-Thinker, a multimodal language model capable of reasoning over textual, visual, and auditory cues. We construct an instruction-tuning dataset with explicit object-aware think-answer chains for Ref-Thinker fine-tuning. The object description inferred by Ref-Thinker is used as an explicit prompt for Grounding-DINO and SAM2, which perform grounding and segmentation without relying on pixel-level supervision. Additionally, we introduce R2-AVSBench, a new benchmark with linguistically diverse and reasoning-intensive references for better evaluating model generalization. Our approach achieves state-of-the-art results on both standard Ref-AVSBench and proposed R2-AVSBench.