AAAI 2026 Papers

Spatio-Temporal Video Grounding (STVG) aims to localize a spatio-temporal tube that corresponds to a given language query in an untrimmed video. This is a challenging task since it involves complex vision-language understanding and spatiotemporal reasoning. Recent works have explored weakly-supervised setting in STVG to eliminate reliance on fine-grained annotations like bounding boxes or temporal stamps. However, they typically follow a simple late-fusion manner, which generates tubes independent of the text description, often resulting in failed target identification and inconsistent target tracking. To address this limitation, we propose a Tube-conditioned Reconstruction with Mutual Constraints (TubeRMC) framework that generates text-conditioned candidate tubes with pre-trained visual grounding models and further refine them via tube-conditioned reconstruction with spatio-temporal constraints. Specifically, we design three reconstruction strategies from temporal, spatial, and spatio-temporal perspectives to comprehensively capture rich tube-text correspondences. Each strategy is equipped with a Tube-conditioned Reconstructor, utilizing spatio-temporal tubes as condition to reconstruct the key clues in the query. We further introduce mutual constraints between spatial and temporal proposals to enhance their quality for reconstruction. TubeRMC outperforms existing methods on two public benchmarks VidSTG and HCSTVG. Further visualization shows that TubeRMC effectively mitigates both target identification errors and inconsistent tracking.

Cross-Video Reasoning (CVR) presents a significant challenge in video understanding, which requires simultaneous understanding of multiple videos to aggregate and compare information across groups of videos. Most existing video understanding benchmarks focus on single-video analysis, failing to assess the ability of multimodal large language models (MLLMs) to simultaneously reason over various videos. Recent benchmarks evaluate MLLMs' capabilities on multi-view videos that capture different perspectives of the same scene. However, their limited tasks hinder a thorough assessment of MLLMs in diverse real-world CVR scenarios. To this end, we introduce CrossVid, the first benchmark designed to comprehensively evaluate MLLMs' spatial-temporal reasoning ability in cross-video contexts. Firstly, CrossVid encompasses a wide spectrum of hierarchical tasks, comprising four high-level dimensions and ten specific tasks, thereby closely reflecting the complex and varied nature of real-world video understanding. Secondly, CrossVid provides 5,331 videos, along with 9,015 challenging question-answering pairs, spanning single-choice, multiple-choice, and open-ended question formats. Through extensive experiments on various open-source and closed-source MLLMs, we observe that Gemini-2.5-Pro performs best on CrossVid, achieving an average accuracy of 50.4%. Notably, our in-depth case study demonstrates that most current MLLMs struggle with CVR tasks, primarily due to their inability to integrate or compare evidence distributed across multiple videos for reasoning. These insights highlight the potential of CrossVid to guide future advancements in enhancing MLLMs’ CVR capabilities.

We present MIRA (Multimodal Interventional RAdiology evaluation), a comprehensive benchmark for evaluating large multimodal models in expert-level interventional radiology tasks requiring specialized domain knowledge and advanced visual reasoning capabilities. Unlike existing medical benchmarks that primarily provide binary labels without contextual depth, MIRA offers diverse question formats, including open-ended, closed-ended, single-choice, and multiple-choice categories, each accompanied by detailed expert-validated explanations. The benchmark incorporates approximately 184K high-quality medical images spanning multiple imaging modalities with 1.2M meticulously generated question-answer pairs across various anatomical regions. These pairs were created through a sophisticated cascade methodology involving expert interventional radiologists at both the data collection and validation stages. Our comprehensive evaluation, encompassing zero-shot testing and fine-tuning experiments of large multimodal models, revealing significant performance gaps between AI systems and human specialists. Fine-tuning experiments demonstrate substantial improvements, with models achieving up to 0.80 accuracy on single-choice questions. MIRA establishes a challenging benchmark that suggests promising directions for developing specialized clinical AI systems for interventional radiology.

Infrared and visible image fusion aims to integrate complementary multi-modal information into a single fused result. However, existing methods 1) fail to account for the degradation visible images under adverse weather conditions, thereby compromising fusion performance; and 2) rely on fixed network architectures, limiting their adaptability to diverse degradation scenarios. To address these issues, we propose a one-stop degradation-aware image fusion framework for multi-degradation scenarios driven by a large language model (MdaIF). Given the distinct scattering characteristics of different degradation scenarios (e.g., haze, rain, and snow) in atmospheric transmission, a mixture-of-experts (MoE) system is introduced to tackle image fusion across multiple degradation scenarios. To adaptively extract diverse weather-aware degradation knowledge and scene feature representations, collectively referred to as the semantic prior, we employ a pre-trained vision-language model (VLM) in our framework. Guided by the semantic prior, we propose degradation-aware channel attention module (DCAM), which employ degradation prototype decomposition to facilitate multi-modal feature interaction in channel domain. In addition, to achieve effective expert routing, the semantic prior and channel-domain modulated features are utilized to guide the MoE, enabling robust image fusion in complex degradation scenarios. Extensive experiments validate the effectiveness of our MdaIF, demonstrating superior performance over SOTA methods.

The remarkable realism of images generated by diffusion models poses critical detection challenges. Current methods utilize reconstruction error as a discriminative feature, exploiting the observation that real images exhibit higher reconstruction errors when processed through diffusion models. However, these approaches require costly reconstruction computations and depend on specific diffusion models, making their performance highly model-dependent. We identify a fundamental difference: real images are more difficult to fit with Gaussian distributions compared to synthetic ones. In this paper, we propose Forgery Identification via Noise Disturbance (FIND), a novel method that requires only a simple binary classifier. It eliminates reconstruction by directly targeting the core distributional difference between real and synthetic images. Our key operation is to add Gaussian noise to real images during training and label these noisy versions as synthetic. This step allows the classifier to focus on the statistical patterns that distinguish real from synthetic images. We theoretically prove that the noise-augmented real images resemble diffusion-generated images in their ease of Gaussian fitting. Furthermore, simply by adding noise, they still retain visual similarity to the original images, highlighting the most discriminative distribution-related features. The proposed FIND improves performance by 11.7% on the GenImage benchmark while running 126x faster than existing methods. By removing the need for auxiliary diffusion models and reconstruction, it offers a practical, efficient, and generalizable way to detect diffusion-generated content.

Unsupervised domain adaptive pose estimation is a fundamental yet challenging task due to the need to transfer from labeled synthetic data to unlabeled real data. Nevertheless, the underlying pose semantics, which are governed by spatial structure, remain largely consistent across domains. This observation motivates the use of vision-language models, which provide domain-invariant representations that align well with high-level semantic concepts. Motivated by this, we propose CLIP2Pose, a novel framework that leverages the semantic robustness of frozen CLIP encoders to facilitate cross-domain generalization. We first introduce a semantic-driven prompt mechanism that encodes structural priors, domain-specific appearance, and instance-level context into the image representation. This guides the model to focus on semantically meaningful and structurally relevant features. Next, we propose a semantic modulation module that adaptively refines visual features by conditioning them on prompt-derived embeddings, enhancing alignment between semantics and visual patterns. To further bridge the modality and domain gaps, we design a directional alignment loss that encourages consistent structural reasoning across both vision and language representations. Extensive experiments on domain adaptive human body and hand pose benchmarks show that CLIP2Pose achieves state-of-the-art performance.

Representation learning is fundamental to modern machine learning, powering applications such as text retrieval and multimodal understanding. However, learning robust and generalizable representations remains challenging. While prior work has demonstrated that active noise injection, a form of data augmentation, can enhance encoding performance, most existing methods rely on heuristic or static noise, overlooking the dynamic nature of feature distributions during training. In this work, we systematically study the role of noise in representation learning from both gradient-based and feature distribution perspectives, using InfoNCE loss as a representative example. Focusing on multimodal representation learning, we propose FANoise, a novel feature-adaptive noise injection strategy. By leveraging the dynamics of contrastive learning, FANoise effectively mitigates the negative impacts of noise while preserving its benefits. Under this theoretically grounded framework, comprehensive experiments demonstrate that FANoise consistently improves overall performance on multimodal tasks across various base VLM models.

We present ReCAD, a reinforcement learning (RL) framework that bootstraps pretrained large models (PLMs) to generate precise parametric computer-aided design (CAD) models from multimodal inputs by leveraging their inherent generative capabilities. With just access to simple functional interfaces (e.g., point coordinates), our approach enables the emergence of complex CAD operations (e.g., pattern replication and mirror). This stands in contrast to previous methods, which typically rely on knowledge injected through supervised fine-tuning (SFT), offer limited support for editability, and fail to exploit the strong generative priors of PLMs. Specifically, the ReCAD framework begins by fine-tuning vision-language models (VLMs) to equip them with basic CAD model generation capabilities, where we rewrite CAD scripts into parameterized code that is leveraged to generate accurate textual descriptions for supervision. Then, we propose a novel RL strategy that incorporates parameterized code as guidance to enhance the model’s reasoning on challenging questions. Furthermore, we employ a hierarchical primitive learning process to progressively teach structured and compositional skills under a unified reward function that ensures both geometric accuracy and semantic fidelity. ReCAD sets a new state-of-the-art in both text-to-CAD and image-to-CAD tasks, significantly improving geometric accuracy across in-distribution and out-of-distribution settings. In the image-to-CAD task, for instance, it reduces the mean Chamfer Distance from 73.47 to 29.61 (in-distribution) and from 272.06 to 80.23 (out-of-distribution), outperforming existing baselines by a substantial margin.

Open-vocabulary semantic segmentation (OVSS) employs pixel-level vision-language alignment to associate category-related prompts with corresponding pixels. A key challenge is enhancing the multimodal dense prediction capability, specifically this pixel-level multimodal alignment. Although existing methods achieve promising results by leveraging CLIP’s vision-language alignment, they rarely investigate the performance boundaries of CLIP for dense prediction from an interpretability mechanisms perspective. In this work, we systematically investigate CLIP's internal mechanisms and identify a critical phenomenon: analogous to human distraction, CLIP diverts significant attention resources from target regions to irrelevant tokens. Our analysis reveals that these tokens arise from dimension-specific over-activation; filtering them enhances CLIP's dense prediction performance. Consequently, we propose Refocusing CLIP (RF-CLIP), a training-free approach that emulates human distraction-refocusing behavior to redirect attention from distraction tokens back to target regions, thereby refining CLIP's multimodal alignment granularity. Our method achieves SOTA performance on eight benchmarks while maintaining high inference efficiency.

Unlike traditional visual segmentation, audio-visual segmentation (AVS) requires the model not only to identify and segment objects but also to determine whether they are sound sources. Recent AVS approaches have achieved impressive performance on standard benchmarks. Yet, an important question remains: Do these models genuinely integrate audio-visual cues to segment sounding objects? Our study reveals a fundamental bias in current methods: they tend to generate segmentation masks based predominantly on visual salience, irrespective of the audio context, resulting in unreliable predictions when sounds are absent or irrelevant. To address this challenge, we introduce AVSBench-Robust, a comprehensive benchmark incorporating diverse negative audio scenarios, including silence, noise, and off-screen sounds. We also propose a simple yet effective approach combining balanced training with negative samples and classifier-guided similarity learning. Our extensive experiments show that while state-of-the-art AVS methods consistently fail under negative audio conditions, our approach achieves remarkable improvements in both standard metrics and robustness measures, maintaining near-perfect false positive rates while preserving high-quality segmentation performance.

Existing plug-and-play image restoration methods typically employ off-the-shelf Gaussian denoisers as proximal operators within classical optimization frameworks based on variable splitting. Recently, denoisers induced by generative priors have been successfully integrated into regularized optimization methods for image restoration under Gaussian noise. However, their application to non-Gaussian noise--such as impulse noise--remains largely unexplored. In this paper, we propose a plug-and-play image restoration framework based on generative diffusion priors for robust removal of general noise types, including impulse noise. Within the maximum a posteriori (MAP) estimation framework, the data fidelity term is adapted to the specific noise model. Departing from the conventional least-squares loss used for Gaussian noise, we introduce a generalized Gaussian scale mixture-based loss, which approximates a wide range of noise distributions and leads to an ℓq-norm fidelity term. This optimization problem is addressed using an iteratively reweighted least squares (IRLS) approach, wherein the proximal step involving the generative prior is efficiently performed via a diffusion-based denoiser. Experimental results on benchmark datasets demonstrate that the proposed method effectively removes non-Gaussian impulse noise and achieves superior restoration performance.

Regularized optimization has been a classical approach to solving imaging inverse problems, where the regularization term enforces desirable properties of the unknown image. Recently, the integration of flow matching generative models into image restoration has garnered significant attention, owing to their powerful prior modeling capabilities. In this work, we incorporate such generative priors into a Plug-and-Play (PnP) framework based on proximal splitting, where the proximal operator associated with the regularizer is replaced by a time-dependent denoiser derived from the generative model. While existing PnP methods have achieved notable success in inverse problems with smooth squared ℓ2 data fidelity--typically associated with Gaussian noise--their applicability to more general data fidelity terms remains underexplored. To address this, we propose a general and efficient PnP algorithm inspired by the primal-dual hybrid gradient (PDHG) method. Our approach is computationally efficient, memory-friendly, and accommodates a wide range of fidelity terms. In particular, it supports both ℓ1 and ℓ2 norm-based losses, enabling robustness to non-Gaussian noise types such as Poisson and impulse noise. We validate our method on several image restoration tasks, including denoising, super-resolution, deblurring, and inpainting, and demonstrate that ℓ1 and ℓ2 fidelity terms outperform the conventional squared ℓ2 loss in the presence of non-Gaussian noise.

Vector quantization has emerged as a powerful tool in large-scale multimodal models, unifying heterogeneous representations through discrete token encoding. However, its effectiveness hinges on robust codebook design. Current prototype-based approaches relying on trainable vectors or clustered centroids fall short in representativeness and interpretability, even as multimodal alignment demonstrates its promise in vision-language models. To address these limitations, we propose a simple multimodal prompting-driven quantization framework for point cloud analysis. Our methodology is built upon two core insights: 1) Text embeddings from pre-trained models inherently encode visual semantics through many-to-one contrastive alignment, naturally serving as robust prototype priors; and 2) Multimodal prompts enable adaptive refinement of these prototypes, effectively mitigating vision-language semantic gaps. The framework introduces a dual-constrained quantization space, enforced by compactness and separation regularization, which seamlessly integrates visual and prototype features, resulting in hybrid representations that jointly encode geometric and semantic information. Furthermore, we employ Gumbel-Softmax relaxation to achieve differentiable discretization while maintaining quantization sparsity. Extensive experiments on the ModelNet40 and ScanObjectNN datasets clearly demonstrate the superior effectiveness of the proposed method.

Person search is a challenging computer vision task that aims to simultaneously detect and re-identify individuals from uncropped gallery images. However, most existing approaches are limited by restricted receptive fields, leading to distorted local feature representations under occlusions or complex poses. Additionally, scale variations hinder model generalization in real-world scenarios. To address these limitations, we introduce a novel E-Bike Rider Search (EBRS) dataset, which comprises 27,501 images capturing 963 distinct IDs across 8 camera views at a large urban intersection in a Chinese city. Furthermore, we propose a Context-aware Dynamic Contrastive Learning (CDCL) framework that dynamically adjusts convolutional weights and performs hard sample mining based on contextual cues, thereby improving discriminative capability for both local details and global features. Extensive experiments show our method achieves state-of-the-art performance on CUHK-SYSU and PRW benchmarks, with competitive results on the challenging EBRS dataset, demonstrating its effectiveness.

Multi-modal object Re-identification (ReID) aims to retrieve individuals by leveraging complementary information from different modalities. Recent CLIP-based approaches show promising results, but they usually employ prompt-based or hybrid prompt-adapter tuning and still face the problems of heterogeneous domain gap, fine-grained identity discrimination and noise instance interference. To address these problems, we introduce a novel Parameter-Efficient Fine-Tuning framework with Bag-of-Adapters (PEFT-BoA) based on the pre-trained CLIP's vision encoder for multi-modal object ReID. Specifically, we first propose a Domain-specific Patch Adapter (DPA) designed to bridge the visual feature gap between pre-trained and fine-tuned models at the local patch level. Meanwhile, we propose a Task-specific Class Adapter (TCA) enhance the fine-grained identity discrimination ability by optimizing global class token. Finally, we propose an Instance-specific Fusion Adapter (IFA) dynamically selects and combines only the most useful features across different modalities for each instance. Our PEFT-BoA achieves the better performance on multi-modal object re-identification benchmarks, while maintaining fewer trainable parameters (6.62M) and a higher training throughput (246.2fps).

Due to the diversity of brain anatomy and the scarcity of annotated data, supervised anomaly detection for brain MRI remains challenging, driving the development of unsupervised anomaly detection (UAD) approaches. Current UAD methods typically utilize synthetically generated noise perturbations on healthy MRIs to train generative models for normal anatomy reconstruction, enabling anomaly detection via residual maps. However, such simulated anomalies lack the biophysical fidelity and morphological complexity characteristic of true clinical lesions. To advance UAD in brain MRI, we conduct the first systematic frequency-domain analysis of pathological signatures, revealing two key properties: (1) anomalies exhibit unique frequency patterns distinguishable from normal anatomy, and (2) low-frequency signals maintain consistent representations across healthy scans. These insights motivate our Frequency-Decomposition Preprocessing (FDP) framework—the first UAD method to leverage frequency-domain reconstruction for simultaneous pathology suppression and anatomical preservation. FDP can integrate seamlessly with existing anomaly simulation techniques, consistently enhancing detection performance across diverse architectures while maintaining diagnostic fidelity. Experimental results demonstrate that FDP consistently improves anomaly detection performance when integrated with existing methods. Notably, FDP achieves a 17.63% increase in DICE score with LDM while maintaining robust improvements across multiple baselines.

RGB-Thermal (RGBT) tracking aims to exploit visible and thermal infrared modalities for robust all-weather object tracking. However, existing RGBT trackers struggle to resolve modality discrepancies, which poses great challenges for robust feature representation. This limitation hinders effective cross-modal information propagation and fusion, which significantly reduces the tracking accuracy. To address this limitation, we propose a novel Contextual Aggregation with Deformable Alignment framework called CADTrack for RGBT Tracking. To be specific, we first deploy the Mamba-based Feature Interaction (MFI) that establishes efficient feature interaction via state space models. This interaction module can operate with linear complexity, reducing computational cost and improving feature discrimination. Then, we propose the Contextual Aggregation Module (CAM) that dynamically activates backbone layers through sparse gating based on the Mixture-of-Experts (MoE). This module can encode complementary contextual information from cross-layer features. Finally, we propose the Deformable Alignment Module (DAM) to integrate deformable sampling and temporal propagation, mitigating spatial misalignment and localization drift. With the above components, our CADTrack achieves robust and accurate tracking in complex scenarios. Extensive experiments on five RGBT tracking benchmarks verify the effectiveness of our proposed method.

As a challenging vision-language task, Zero-Shot Composed Image Retrieval (ZS-CIR) is designed to retrieve target images using bi-modal (image+text) queries. Typical ZS-CIR methods employ an inversion network to generate pseudo-word tokens that effectively represent the input semantics. However, the inversion-based methods suffer from two inherent issues: First, the task discrepancy exists because inversion training and CIR inference involve different objectives. Second, the modality discrepancy arises from the input feature distribution mismatch between training and inference. To this end, we propose a lightweight post-hoc framework, consisting of two components: (1) A new text-anchored triplet construction pipeline leverages a large language model (LLM) to transform a standard image-text dataset into a triplet dataset, where a textual description serves as the target of each triplet. (2) The MoTa-Adapter, a novel parameter-efficient fine-tuning method, adapts the dual encoder to the CIR task using our constructed triplet data. Specifically, on the text side, multiple sets of learnable task prompts are integrated via a Mixture-of-Experts (MoE) layer to capture task-specific priors and handle different types of modifications. On the image side, MoTa-Adapter modulates the inversion network's input to better match the downstream text encoder. In addition, an entropy-based optimization strategy is proposed to assign greater weight to challenging samples, thus improving adaptation efficiency. Experiments show that, with the incorporation of our proposed components, inversion-based methods achieve significant improvements, reaching state-of-the-art performance across four widely-used benchmarks.

Generating interaction-centric videos, such as those depicting humans or robots interacting with objects, is crucial for embodied intelligence, as they provide rich and diverse visual priors for robot learning, manipulation policy training, and affordance reasoning. However, existing methods often struggle to model such complex and dynamic interactions. While recent studies show that masks can serve as effective control signals and enhance generation quality, obtaining dense and precise mask annotations remains a major challenge for real-world use. To overcome this limitation, we introduce Mask2IV, a novel framework specifically designed for interaction-centric video generation. It adopts a decoupled two-stage pipeline that first predicts plausible motion trajectories for both actor and object, then generates a video conditioned on these trajectories. This design eliminates the need for dense mask inputs from users while preserving the flexibility to manipulate the interaction process. Furthermore, Mask2IV supports versatile and intuitive control, allowing users to specify the target object of interaction and guide the motion trajectory through action descriptions or spatial position cues. To support systematic training and evaluation, we curate two benchmarks covering diverse action and object categories across both human-object interaction and robotic manipulation scenarios. Extensive experiments demonstrate that our method achieves superior visual realism and controllability compared to existing baselines.