AAAI 2026 Papers

Although Vision Language Models (VLMs) have excelled at image and video understanding, applying them to hour-long videos is held back by two interrelated challenges: exorbitant computational expense and a qualitative breakdown in long-term temporal reasoning. Thus, models tend to generate answers based on speculation instead of solid visual facts, causing both factually incorrect and plausible hallucinations. This problem is compounded by current benchmarks that, by only emphasizing final answers, lack an effective mechanism to check whether reasoning is substantiated by specific visual evidence. This makes it hard to differentiate between true understanding and pretend comprehension, inhibiting targeted model refinement. To address these interrelated challenges of model fragility and evaluation weakness, we adopt a twofold strategy. First, we present EV²-Bench, a large-scale benchmark that breaks new ground by an evaluation paradigm built upon spatio-temporal visual evidence, forcing models to justify answers with checkable hints. Second, we put forward DynamicSelect, an adaptive token compression system that efficiently condenses salient information by a dynamic semantic selector and a hierarchical compression strategy. Comprehensive experiments demonstrate that DynamicSelect significantly outperforms the baselines on EV²-Bench as well as other public benchmarks. Our study offers not only a more effective approach to long-video understanding but also a more stringent evaluation paradigm, indicating the way toward more robust models.

Generating high-quality physically based rendering (PBR) materials is important to achieve realistic rendering in the downstream tasks, yet it remains challenging due to the intertwined effects of materials and lighting. While existing methods have made breakthroughs by incorporating material decomposition in the 3D generation pipeline, they tend to bake highlights into albedo and ignore spatially varying properties of metallicity and roughness. In this work, we present PBR3DGen, a two-stage mesh generation method with high-quality PBR materials that integrates the novel multi-view PBR material estimation model and a 3D PBR mesh reconstruction model. Specifically, PBR3DGen leverages vision language models (VLM) to guide multi-view diffusion, precisely capturing the spatial distribution and inherent attributes of reflective-metalness material. Additionally, we incorporate view-dependent illumination-aware conditions as pixel-aware priors to enhance spatially varying material properties. Furthermore, our reconstruction model reconstructs high-quality mesh with PBR materials. Experimental results demonstrate that PBR3DGen significantly outperforms existing methods, achieving new state-of-the-art results for PBR estimation and mesh generation.

Many methods have demonstrated promising results in zero-shot anomaly detection (ZSAD) by incorporating prompt learning (PL) to fine-tune Vision-Language Models. However, the prompt learners proposed in recent studies remain relatively simple, such as learnable textual and visual prompts. Relying solely on the current PL paradigm restricts the ability to generate more precise prompts, thereby hindering improved ZSAD performance. To mitigate this issue, this paper proposes a high-order-aware prompt learning framework, termed HiPL, which facilitates the detection of unseen anomalies through generating prompts fortified by hypergraphs. Specifically, HiPL models high-order correlations among patches through a dynamically constructed hypergraph structure. Then we leverage a hypergraph semantic convolution to capture potential collaborative information by propagating high-order correlations by hyperedges. Meanwhile, HiPL introduces a Mixture-of-Experts prompt learner (MoEPLer), where the experts within MoEPLer can generate multiple distinct prompts based on the modeled high-order correlations. Then, the final high-order-aware textual prompts can be formed by synthetically considering each expert's prompt by gating weights. This enables a comprehensive understanding of potential anomalous patterns, thereby facilitating ZSAD performance. Large-scale experiments conducted on 12 datasets, spanning natural, industrial, and medical domains, demonstrate the validity of proposed HiPL.

Cross-modal Knowledge Distillation has demonstrated promising performance on paired modalities with strong semantic connections, referred to as Symmetric Cross-modal Knowledge Distillation (SCKD). However, implementing SCKD becomes exceedingly constrained in real-world scenarios due to the limited availability of paired modalities. To this end, we investigate a general and effective knowledge learning concept under weak semantic consistency, dubbed Asymmetric Cross-modal Knowledge Distillation (ACKD), aiming to bridge modalities with limited semantic overlap. Nevertheless, the shift from strong to weak semantic consistency improves flexibility but exacerbates challenges in knowledge transmission costs, which we rigorously verified based on optimal transport theory. To mitigate the issue, we further propose a framework, namely SemBridge, integrating a Student-Friendly Matching module and a Semantic-aware Knowledge Alignment module. The former leverages self-supervised learning to acquire semantic-based knowledge and provide personalized instruction for each student sample by dynamically selecting the relevant teacher samples. The latter seeks the optimal transport path by employing Lagrangian optimization. To facilitate the research, we curate a benchmark dataset derived from two modalities, namely Multi-Spectral (MS) and asymmetric RGB images, tailored for remote sensing scene classification. Comprehensive experiments exhibit that our framework achieves state-of-the-art performance compared with 7 existing approaches on 6 different model architectures across various datasets.

Enabling intuitive, language-driven interaction with surgical scenes is a critical step toward intelligent operating rooms and autonomous surgical robotic assistance. However, the task of referring segmentation, localizing surgical instruments based on natural language descriptions, remains underexplored in surgical videos, with existing approaches struggling to generalize due to reliance on static visual cues and predefined instrument names. In this work, we introduce SurgRef, a novel motion-guided framework that grounds free-form language expressions in instrument motion, capturing how tools move and interact across time, rather than what they look like. This allows models to understand and segment instruments even under occlusion, ambiguity, or unfamiliar terminology. To train and evaluate SurgRef, we present Ref-IMotion, a diverse, multi-institutional video dataset with dense spatiotemporal masks and rich motion-centric expressions. SurgRef achieves state-of-the-art accuracy and generalization across surgical procedures, setting a new benchmark for robust, language-driven surgical video segmentation.

Masked autoencoders (MAE) have become a dominant paradigm in 3D representation learning, setting new performance benchmarks across various downstream tasks. Existing methods with fixed mask ratios neglect multi-level representational correlations and intrinsic geometric structures, while relying on point-wise reconstruction assumptions that conflict with the diversity of point cloud. To address these issues, we propose a 3D representation learning method, termed Point-SRA, which aligns representations through self-distillation and probabilistic modeling. Specifically, we assign different masking ratios to the MAE to capture complementary geometric and semantic information, while the MeanFlow Transformer (MFT) leverages cross-modal conditional embeddings to enable diverse probabilistic reconstruction. Our analysis further reveals that representations at different time steps in MFT also exhibit complementarity. Therefore, a Dual Self-Representation Alignment mechanism is proposed at both the MAE and MFT levels. Finally, we design a Flow-Conditioned Fine-Tuning Architecture to fully exploit the point cloud distribution learned via MeanFlow. Point-SRA outperforms Point-MAE by 5.37% on ScanObjectNN. On intracranial aneurysm segmentation, it reaches 96.07% mean IoU for arteries and 86.87% for aneurysms. For 3D object detection, Point-SRA achieves 47.3% AP@50, surpassing MaskPoint by 5.12%.

Scene text segmentation is a critical preprocessing step in various text-based applications. Specialist text segmentation methods, often relying on a detect-then-segment paradigm, tend to exhibit reduced robustness and can lead to cascading errors. The introduction of the Segment Anything Model (SAM) has revolutionized general segmentation by leveraging vision foundation models. However, SAM still falls short when applied to domain-specific tasks such as scene text segmentation. To bridge this gap between SAM and specialized scene text segmentation approaches, we propose ST-SAM (Scene Text SAM), a parameter-efficient fine-tuning framework tailored to adapt SAM for high-quality scene text segmentation without relying on explicit text detection. ST-SAM incorporates a multimodal prompting mechanism: a lightweight visual encoder generates multi-scale spatial features to provide precise visual context; and textual prompts generated by a large language model offer high-level semantic guidance. We demonstrate the advantages of the proposed ST-SAM as follows: (1) ST-SAM achieves new state-of-the-art performance on multiple scene text segmentation benchmarks, including 85.30% fgIoU on Total-Text and 91.03% fgIoU on TextSeg, outperforming both specialist and generalist models. (2) ST-SAM enables effective domain adaptation by flexibly adapting the general SAM architecture to the domain of scene text. (3) By discarding the detect-then-segment pipeline, ST-SAM simplifies the inference process while still achieving robust performance on complex text cases.

Recent diffusion-based Single-image 3D portrait generation methods typically employ 2D diffusion models to provide multi-view knowledge, which is then distilled into 3D representations. However, these methods usually struggle to produce high-fidelity 3D models, frequently yielding excessively blurred textures. We attribute this issue to the insufficient consideration of cross-view consistency during the diffusion process, resulting in significant disparities between different views and ultimately leading to blurred 3D representations. In this paper, we address this issue by comprehensively exploiting multi-view priors in both the conditioning and diffusion procedures to produce consistent, detail-rich portraits. From the conditioning standpoint, we propose a Hybrid Priors Diffusion model, which explicitly and implicitly incorporates multi-view priors as conditions to enhance the status consistency of the generated multi-view portraits. From the diffusion perspective, considering the significant impact of the diffusion noise distribution on detailed texture generation, we propose a Multi-View Noise Resampling Strategy integrated within the optimization process leveraging cross-view priors to enhance representation consistency. Extensive experiments show that our method produces 3D portraits with accurate geometry and rich details from a single image.

Detecting mirror regions in RGB videos is essential for scene understanding in applications such as scene reconstruction and robotic navigation. Existing video mirror detectors typically rely on cues like inside-outside mirror correspondences and 2D motion inconsistencies. However, these methods often yield noisy or incomplete predictions when confronted with complex real-world video scenes, especially in areas with occlusion or limited visual features and motions. We observe that human perceive and navigate 3D occluded environments with remarkable ease, owing to Motion-in-Depth (MiD) perception. MiD integrates information from visual appearance (image colors and textures), the way objects move around us in 3D space (3D motions), and their relative distance from us (depth) to determine if something is approaching or receding and to support navigation. Motivated by this neuroscience mechanism, we introduce MiD-VMD, the first approach to explicitly model MiD for video mirror detection. MiD-VMD jointly utilizes contrastive 3D motion, depth, and image features through two novel modules based on a combinational QKV transformer architecture. The Motion-in-Depth Attention Learning (MiD-AL) module captures complementary relationships across these modalities with combinatorial attention and enforces a compact encoding to represent global 3D transformations, resulting in more accurate mirror detection and reduced motion artifacts. The Motion-in-Depth Boundary Detection (MiD-BD) module further sharpens mirror boundaries by leveraging cross-modal attention on 3D motion and depth features. Extensive experiments show that MiD-VMD outperforms current SOTAs.

Storytelling video generation (SVG) aims to produce coherent and visually rich multi-scene videos that follow a structured narrative. Existing methods primarily employ LLM for high-level planning to decompose a story into scene-level descriptions, which are then independently generated and stitched together. However, these approaches struggle with generating high-quality videos aligned with the complex single-scene description, as visualizing such complex description involves coherent composition of multiple objects/events, complex motion synthesis and character customization with sequential motions. To address these challenges, we propose DREAMRUNNER, a novel story-to-video generation method: First, we structure the input script using a large language model (LLM) to facilitate both coarse-grained scene planning as well as fine-grained object-level layout planning. Next, DREAMRUNNER presents retrieval-augmented test-time adaptation to capture target motion priors for objects in each scene, supporting diverse motion customization based on retrieved videos, thus facilitating the generation of new videos with complex, scripted motions. Lastly, we propose a novel spatial-temporal region-based 3D attention and prior injection module SR3AI for fine-grained object-motion binding and frame-by-frame spatial-temporal semantic control. We compare DREAMRUNNER with various SVG baselines, demonstrating state-of-the-art performance in character consistency, text alignment, and smooth transitions. Additionally, DREAMRUNNER exhibits strong fine-grained condition-following ability in compositional text-to-video generation, significantly outperforming baselines on T2V-ComBench. Finally, we demonstrate DREAMRUNNER’s ability to generate multi-character interactions with qualitative examples.

Modern gaze estimation models can accurately predict human gaze from facial images. However, due to privacy concerns and intricate data collection procedures, gaze estimation datasets are typically smaller and less diverse compared to those for other vision tasks, which directly leads to poor generalization in gaze estimation models. Common solutions, such as domain adaptation models, require additional domain-specific data, yet such data is often difficult to obtain due to privacy restrictions. Meanwhile, domain generalization models suffer from limited performance due to insufficient training data. To address these fundamental challenges---privacy and data diversity---we explore privacy-preserving gaze data generation schemes and propose a novel data-driven generalization solution. Specifically, we develop two diffusion-based generative models, DDPM-Gaze and LDM-Gaze, for synthesizing gaze data. We demonstrate that synthetic data can significantly improve generalization performance when simply used with fine-tuning-based methods. Furthermore, we introduce the Domain Stability Adaptation (DSA) framework, a simple yet effective domain generalization approach that enhances model robustness by increasing the domain uncertainty of input samples while reducing prediction uncertainty. Extensive experiments validate the effectiveness of our synthetic data and demonstrate the superiority of our data-driven generalization solution.

Large Vision-Language Models (LVLMs) enhance performance on vision-language tasks by integrating visual features from pre-trained vision encoders into large language models (LLMs). However, the large number of visual tokens introduces significant computational overhead. Existing token pruning methods either perform global selection via [CLS]-based attention in the vision encode or prune within LLM decoding layers. These approaches face two key challenges: (1) [CLS]-based attention primarily focuses on visually salient regions across the entire image, often overlooking semantically important tokens essential for reasoning; and (2) strong positional bias in the shallow decoder layers causes the model to favor later-positioned tokens, while neglecting earlier ones that may carry critical reasoning cues. To address these issues, we propose PosPrune, a training-free, two-stage visual token pruning framework. At the vision encoder, we introduce an Asymmetric Region-aware Pruning (ARP) strategy that retains more tokens in semantically rich regions while discarding more tokens from semantically less informative regions, thus preserving spatial diversity and task-relevant details. In the LLM decoding stage, we find that the positional bias in shallow layers is primarily driven by model architecture rather than task semantics. Based on this insight, we propose a novel Positional Bias Correction (PBC) mechanism to mitigate this bias. To further reduce redundancy, we apply Maximal Marginal Relevance (MMR) to select tokens that best balance textual relevance and diversity. Extensive experiments on various LVLMs and benchmarks demonstrate the general effectiveness of our approach. Notably, when applied to LLaVA-1.5-7B, PosPrune achieves a reduction of 85% in FLOPs while preserving 98.5% of the original performance.

Motion estimation in degraded scenes has long been a significant challenge, primarily attributed to substantial scene variations and insufficient training data. Existing approaches typically address this limitation by incorporating additional training strategies or modifying network architectures within conventional frameworks. However, these solutions not only require cumbersome training procedures or additional modal inputs, but also lack generalization capabilities. To address this problem, we propose a unified optical flow estimation framework specifically designed for degraded scenes. In this work, we employ large-scale pre-trained optical flow foundation models as both teacher and student networks. Our objective is to compensate for feature incompleteness during image degradation through pre-trained large models. Subsequently, we leverage supervised signals for fine-tuning and introduce an intra-inter frame distillation method to enable the student network to adapt to diverse cross-domain scenarios. Our proposed methodology provides deeper insights into learning style-invariant features from these learnable fine-tuning layers. Extensive experiments demonstrate that our approach achieves superior generalization performance and state-of-the-art results in degraded scenes (including low-light, rain, fog and other conditions) while requiring minimal training resources.

Recent advances in video understanding have been driven by MLLMs. But these MLLMs are good at analyzing short videos, while suffering from difficulties in understanding videos with a longer context. To address this difficulty, several agent paradigms have recently been proposed, using MLLMs as agents for retrieving extra contextual knowledge in a long video. However, most existing agents ignore the key fact that a long video is composed with multiple shots, i.e., to answer the user question from a long video, it is critical to deeply understand its relevant shots like human. Without such insight, these agents often mistakenly find redundant even noisy temporal context, restricting their capacity for long video understanding. To fill this gap, we propose VideoChat-A1, a novel long video agent paradigm. Different from the previous works, our VideoChat-A1 can deeply think with long videos, via a distinct chain-of-shot reasoning paradigm. More specifically, it can progressively select the relevant shots of user question, and look into these shots in a coarse-to-fine partition. By multi-modal reasoning along the shot chain, VideoChat-A1 can effectively mimic step-by-step human thinking process, allowing the interactive discovery of preferable temporal context for thoughtful understanding in long videos. Extensive experiments show that, VideoChat-A1 achieves the state-of-the-art performance on the mainstream long video QA benchmarks, e.g., it achieves 77.0 on VideoMME(w/ subs) and 70.1 on EgoSchema, outperforming its strong baselines (e.g., InternVL2.5-8B and InternVideo2.5-8B), by up to 10.1% and 6.2%. Compared to leading closed-source GPT-4o and Gemini 1.5 Pro, VideoChat-A1 offers competitive accuracy, but only with 7% input frames and 12% inference time on average.

In-context learning (ICL) has emerged as a powerful paradigm for Large Visual Language Models (LVLMs), enabling them to leverage a few examples directly from input contexts. However, the effectiveness of this approach is heavily reliant on the selection of demonstrations, a process that is NP-hard. Traditional strategies, including random, similarity-based sampling and infoscore-based sampling, often lead to inefficiencies or suboptimal performance, struggling to balance both efficiency and effectiveness in demonstration selection. In this paper, we propose a novel demonstration selection framework named Coreset-based Dual Retrieval (CoDR). We show that samples within a diverse subset achieve a higher expected mutual information. To implement this, we introduce a cluster-pruning method to construct a diverse coreset that aligns more effectively with the query while maintaining diversity. Additionally, we develop a dual retrieval mechanism that enhances the selection process by achieving global demonstration selection while preserving efficiency. Experimental results demonstrate that our method significantly improves the ICL performance compared to the existing strategies, providing a robust solution for effective and efficient demonstration selection.

Differentially private (DP) image synthesis enables the generation of realistic images while bounding privacy leakage, facilitating secure data sharing across organizations. However, the Gaussian noise injected during DP training, such as via DP-SGD, often severely degrades synthesis quality by disrupting model convergence. To address this, we introduce RPGen, a novel framework that enhances diffusion models' parameter robustness to mitigate DP noise effects without compromising privacy guarantees. At its core, RPGen employs adversarial model perturbation (AMP) during public pre-training to build resilience against perturbations, but we identify and tackle the critical issue of robustness transferability across domains. RPGen achieves this through a three-step process: (1) A pre-trained classifier infers labels for private images, aggregated into a class distribution noised with Gaussian mechanism for DP, and public samples are selected to match this privatized distribution for domain alignment; (2) The diffusion model is pre-trained on this curated subset with adversarial model perturbation to foster robustness; (3) The model undergoes fine-tuning on private data using DP-SGD. This synergy of robustness augmentation and transferability optimization yields high-fidelity synthesis. Extensive evaluations on ImageNet for pre-training, with CelebA and CIFAR-10 for synthesis, show RPGen outperforming state-of-the-art baselines across epsilon in 1, 5, 10. On average, it achieves 20.18% lower FID and 5.45% higher classification accuracy. Ablations confirm the efficacy of domain curation and modest perturbations, establishing RPGen as a new benchmark for privacy-utility trade-offs in image generation.

Recent advances in Large Visual Language Models (LVLMs) have demonstrated impressive performance across various vision-language tasks by leveraging large-scale image-text pretraining and instruction tuning. However, the security vulnerabilities of LVLMs have become increasingly concerning, particularly their susceptibility to backdoor attacks. Existing backdoor attacks focus on single-target attacks, i.e., targeting a single malicious output associated with a specific trigger. In this work, we uncover multi-target backdoor attacks, where multiple independent triggers corresponding to different attack targets are added in a single pass of training, posing a greater threat to LVLMs in real-world applications. Executing such attacks in LVLMs is challenging since there can be many incorrect trigger-target mappings due to severe feature interference among different triggers. To address this challenge, we propose MTAttack, the first multi-target backdoor attack framework for enforcing accurate multiple trigger-target mappings in LVLMs. The core of MTAttack is a novel optimization method with two constraints, namely Proxy Space Partitioning constraint and Trigger Prototype Anchoring constraint. It jointly optimizes multiple triggers in the latent space, with each trigger independently mapping clean images to a unique proxy class while at the same time guaranteeing their separability. Experiments on popular benchmarks demonstrate a high success rate of MTAttack for multi-target attacks, substantially outperforming existing attack methods. Furthermore, our attack exhibits strong generalizability across datasets and robustness against backdoor defense strategies. These findings highlight the vulnerability of LVLMs to multi-target backdoor attacks and underscore the urgent need for mitigating such threats.

Recent advances in video generation techniques have given rise to an emerging paradigm of generative video coding for Ultra-Low Bitrate (ULB) scenarios by leveraging powerful generative priors. However, most existing methods are limited by domain specificity (e.g., facial or human videos) or excessive dependence on high-level text guidance, which tend to inadequately capture fine-grained motion details, leading to unrealistic or incoherent reconstructions. To address these challenges, we propose Trajectory-Guided Generative Video Coding (dubbed T-GVC), a novel framework that bridges low-level motion tracking with high-level semantic understanding. T-GVC features a semantic-aware sparse motion sampling pipeline that extracts pixel-wise motion as sparse trajectory points based on their semantic importance, significantly reducing the bitrate while preserving critical temporal semantic information. In addition, by integrating trajectory-aligned loss constraints into diffusion processes, we introduce a training-free guidance mechanism in latent space to ensure physically plausible motion patterns without sacrificing the inherent capabilities of generative models. Experimental results demonstrate that T-GVC outperforms both traditional and neural video codecs under ULB conditions. Furthermore, additional experiments confirm that our framework achieves more precise motion control than existing text-guided methods, paving the way for a novel direction of generative video coding guided by geometric motion modeling.

Latent Diffusion Models have become a powerful tool for generating high-fidelity unrestricted adversarial examples. However, the existing methods typically perturb only the initial latent or rely on prompt engineering, which is ill-suited to the iterative nature of the diffusion process, plus optimization instability due to external text prompts and cumulative drift that push the adversarial images off the data manifold. In this paper, we propose a hierarchical attack framework that operates in alignment with the model's generative manifold and leverages intermediate denoising states to maximize attack transferability and visual fidelity. Extensive experiments show that the proposed attack improves adversarial transferability by 10-20% against a diverse set of normally-trained models and achieves over 10.5% higher success rate against adversarially-defended models, while simultaneously enhancing visual quality by 1.0-1.2 FID reduction and 16.7% LPIPS improvements.