CVPR 2026 论文列表

Video Question Answering (VideoQA) has made significant strides by leveraging multimodal learning to align visual and textual modalities. However, current benchmarks overwhelmingly focus on questions answerable through explicit visual content - actions, objects, and events - directly observable within individual frames or short clips. To truly understand videos as humans do, models must go beyond what is directly shown, inferring hidden relationships and contextual cues that are only implied across frames. Current benchmarks fail to capture this essential aspect of video understanding. To address this gap, we introduce VRR-QA, a benchmark for Visual Relational Reasoning Beyond Explicit Cues. We curate our benchmark from creative and cinematic videos such as movies, that deliberately employ storytelling techniques which omit direct depictions of certain events or relations, requiring viewers to infer them. VRR-QA comprises 1K meticulously expert-annotated QA pairs drawn from 1K creative video clips covering 15 genres across 7 decades of content, from both live-action and animated titles. Our extensive evaluations on 14 leading VideoQA models reveals consistent and significant performance degradation, underscoring their reliance on surface-level visual cues and highlighting the difficulty of implicit reasoning. Even the best model substantially underperforms human baselines with only 64% accuracy. Performance variations across models further illustrate the complexity and diversity of the challenges presented by VRR-QA. By releasing both dataset and data collection framework, VRR-QA establishes a rigorous, diverse, and reproducible testbed for advancing VideoQA.

Fine-grained spatiotemporal reasoning on surgical videos is critical, yet the capabilities of Multi-modal Large Language Models (MLLMs) in this domain remain largely unexplored. To bridge this gap, we introduce SurgCoT, a unified benchmark for evaluating chain-of-thought (CoT) reasoning in MLLMs across 7 surgical specialties and 35 diverse procedures. SurgCoT assesses five core reasoning dimensions: Causal Action Ordering, Cue-Action Alignment, Affordance Mapping, Micro-Transition Localization, and Anomaly Onset Tracking, through a structured CoT framework with an intensive annotation protocol (Question-Option-Knowledge-Clue-Answer), where the Knowledge field provides essential background context and Clue provides definitive spatiotemporal evidence. Evaluation of 10 leading MLLMs shows: 1) commercial models outperform open-source and medical-specialized variants; 2) significant gaps exist in surgical CoT reasoning; 3) SurgCoT enables effective evaluation and enhances progressive spatiotemporal reasoning. SurgCoT provides a reproducible testbed to narrow the gap between MLLM capabilities and clinical reasoning demands. Code and data: https://github.com/CVI-SZU/SurgCoT.

Self-supervised 3D occupancy prediction offers a promising solution for understanding complex driving scenes without requiring costly 3D annotations. However, training dense occupancy decoders to capture fine-grained geometry and semantics can demand hundreds of GPU hours, and once trained, such models struggle to adapt to varying voxel resolutions or novel object categories without extensive retraining. To overcome these limitations, we propose a practical and flexible test-time occupancy prediction framework termed TT-Occ. Our method incrementally constructs, optimizes, and voxelizes time-aware 3D Gaussians from raw sensor streams by integrating vision foundation models (VFMs) at runtime. The flexible representation of 3D Gaussians enables voxelization at arbitrary user-specified resolutions, while the strong generalization capability of VFMs supports accurate perception and open-vocabulary recognition without requiring any network training or fine-tuning. To validate the generality and effectiveness of our framework, we present two variants: a LiDAR-based version and a vision-centric version, and conduct extensive experiments on the Occ3D-nuScenes and nuCraft benchmarks under varying voxel resolutions. Experimental results show that TT-Occ significantly outperforms existing computationally expensive pretrained self-supervised counterparts. Code is available at https://github.com/Xian-Bei/TT-Occ.

Classical video quality assessment (VQA) methods generate a numerical score to judge a video's perceived visual fidelity and clarity. Yet, a score fails to describe the video's complex quality dimensions (e.g., noise), restricting its applicability. Benefiting from the human-friendly linguistic output, adapting video large multimodal models (LMMs) to VQA via instruction tuning has the potential to address this issue. The core of the approach lies in the video quality-centric instruction data. Previous explorations mainly focus on the image domain, and their data generation processes heavily rely on human quality annotations and proprietary systems (e.g., GPT-4), limiting data scalability and effectiveness. To address these challenges, we propose the Score-based Instruction Generation (SIG) pipeline. Specifically, SIG first scores multiple quality dimensions of an unlabeled video and maps scores to text-defined levels. It then explicitly incorporates a hierarchical Chain-of-Thought (CoT) to model the correlation between specific dimensions and overall quality, mimicking the human visual system's (HVS) reasoning process. The automated pipeline eliminates the reliance on expert-written quality descriptions and proprietary systems, ensuring data scalability and generation efficiency. To this end, the resulting Score2Instruct (S2I) dataset contains over 320K diverse instruction-response pairs, laying the basis for instruction tuning. Moreover, to advance video LMMs' quality scoring and justification abilities simultaneously, we devise a progressive tuning strategy to fully unleash the power of S2I. Built upon SIG, we further curate a benchmark termed S2I-Bench with 400 open-ended questions to better evaluate the quality justification capacity of video LMMs. Experimental results on the S2I-Bench and existing benchmarks indicate that our method consistently improves quality scoring and justification capabilities across multiple video LMMs. The code and dataset will be available at https://github.com/KeiChiTse/S2I.

Existing generative models, such as diffusion and auto-regressive networks, are inherently static, relying on a fixed set of pretrained parameters to handle all inputs. In contrast, humans flexibly adapt their internal generative representations to each perceptual or imaginative context. Inspired by this capability, we introduce Composer, a new paradigm for adaptive generative modeling based on test-time instance-specific parameter composition. Composer generates input-conditioned parameter adaptations at inference time, which are injected into the pretrained model's weights, enabling per-input specialization without fine-tuning or retraining. Adaptation occurs once prior to multi-step generation, yielding higher-quality, context-aware outputs with minimal computational and memory overhead. Experiments show that Composer substantially improves performance across diverse generative models and use cases, including lightweight/quantized models and test-time scaling. By leveraging input-aware parameter composition, Composer establishes a new paradigm for designing generative models that dynamically adapt to each input, moving beyond static parameterization. The code will be available at https://github.com/tmtuan1307/Composer.

The recent success of reinforcement learning (RL) in large reasoning models has inspired the growing adoption of RL for post-training Multimodal Large Language Models (MLLMs) to enhance their visual reasoning capabilities. Although many studies have reported improved performance, it remains unclear whether RL training truly enables models to learn from visual information. In this work, we propose the Hallucination-as-Cue Framework, an analytical framework designed to investigate the effects of RL-based post-training on multimodal reasoning models from the perspective of model hallucination. Specifically, we introduce hallucination-inductive, modality-specific corruptions that remove or replace essential information required to derive correct answers, thereby forcing the model to reason by hallucination. By applying these corruptions during both training and evaluation, our framework provides a unique perspective for diagnosing RL training dynamics and understanding the intrinsic properties of datasets. Through extensive experiments and analyses across multiple multimodal reasoning benchmarks, we reveal that the role of model hallucination for RL-training is more significant than previously recognized. For instance, we find that RL post-training under purely hallucination-inductive settings can still significantly improve models' reasoning performance, and in some cases even outperform standard training. These findings challenge prevailing assumptions about MLLM reasoning training and motivate the development of more modality-aware RL-based training designs.

Humor, as both a creative human activity and a social binding mechanism, has long posed a major challenge for AI generation. Although producing humor requires complex cognitive reasoning and social understanding, theories of humor suggest that it follows learnable patterns and structures, making it theoretically possible for generative models to acquire them implicitly. In recent years, multimodal humor has become a prevalent form of online communication, especially among Gen Z, highlighting the need for AI systems capable of integrating visual understanding with humorous language generation. However, existing data-driven approaches lack explicit modeling or theoretical grounding of humor, often producing literal descriptions that fail to capture its underlying cognitive mechanisms, resulting in the generated image descriptions that are fluent but lack genuine humor or cognitive depth. To address this limitation, we propose HUMORCHAIN (HUmor-guided Multi-step Orchestrated Reasoning Chain for Image Captioning), a theory-guided multi-stage reasoning framework. It integrates visual semantic parsing, humor- and psychology-based reasoning, and a fine-tuned discriminator for humor evaluation, forming an interpretable and controllable cognitive reasoning chain. To the best of our knowledge, this is the first work to explicitly embed cognitive structures from humor theories into multimodal humor generation, enabling a structured reasoning process from visual understanding to humor creation. Experiments on Meme-Image-No-Text, Oogiri-GO, and OxfordTVG-HIC datasets show that HUMORCHAIN outperforms state-of-the-art baselines in human humor preference, Elo/BT scores, and semantic diversity, demonstrating that theory-driven structured reasoning enables large language models to generate humor aligned with human perception.

Predicting wildfire risk is a reasoning-intensive spatial problem that requires the integration of visual, climatic, and geographic factors to infer continuous risk maps. Existing methods lack the causal reasoning and multimodal understanding required for reliable generalization. We introduce FireScope-Bench, a large-scale dataset and benchmark that couples Sentinel-2 imagery and climate data with expert-defined risk rasters across the USA, and real wildfire events in Europe for cross-continental evaluation. Building on this dataset, we propose FireScope, a VLM-based reasoning-to-generation framework that learns from both reinforcement learning and visual supervision to predict risk rasters with complementary reasoning traces. When trained in the USA and tested in Europe, FireScope achieves substantial performance gains, while expert feedback and automated analysis confirm that its reasoning traces are faithful and semantically meaningful. Our findings demonstrate that reasoning can ground raster prediction models, improving both generalization and interpretability. To our knowledge, this is the first framework to (1) demonstrate that language-based reasoning can improve generalization in visual generation, (2) propose a high-resolution wildfire risk model that can be applied across continents, and (3) enable systematic studies of robust cross-continental generalization for multimodal fire risk models. We believe that FireScope-Bench has the potential to serve as a foundation for advancing reasoning-driven, interpretable and generalizable spatial modeling. Data and source code are available at github.com/insait-institute/FireScope.

Text-aerial person retrieval aims to identify targets in UAV-captured images from eyewitness descriptions, supporting intelligent transportation and public security applications. Compared to ground-view text-image person retrieval, UAV-captured images often suffer from degraded visual information due to drastic variations in viewing angles and flight altitudes, making semantic alignment with textual descriptions very challenging. To address this issue, we propose a novel Cross-modal Fuzzy Alignment Network (CFANet), which quantifies the token-level reliability by fuzzy logic to achieve accurate fine-grained alignment and incorporates ground-view images as a bridge agent to further mitigate the gap between aerial images and text descriptions, for text-aerial person retrieval. In particular, we design the Fuzzy Token Alignment module that employs the fuzzy membership function to dynamically model token-level association strength and suppress the influence of unobservable or noisy tokens. It can alleviate the semantic inconsistencies caused by missing visual cues and significantly enhance the robustness of token-level semantic alignment. Moreover, to further mitigate the gap between aerial images and text descriptions, we design a Context-Aware Dynamic Alignment module to incorporates the ground-view agent as a bridge in text-aerial alignment and adaptively combine direct alignment and agent-assisted alignment to improve the robustness. In addition, we construct a large-scale benchmark dataset called AERI-PEDES by using a chain-of-thought to decompose text generation into attribute parsing, initial captioning, and refinement, thus boosting textual accuracy and semantic consistency. Experiments on AERI-PEDES and TBAPR demonstrate the superiority of our method. The code and dataset will be publicly released.

Multimodal Large Reasoning Models (MLRMs) exhibit remarkable performance on complex tasks by incorporating explicit multi-step reasoning. However, this capability also introduces new security vulnerabilities. Existing jailbreak studies largely overlook Cognitive-level weaknesses embedded in the reasoning process itself. In this work, we uncover a critical cognitive bias in MLRMs: the anchoring effect, where safety judgments are disproportionately influenced by the first piece of information received--the anchor. Building on this finding, we propose the Reasoning-chain Anchoring Attack (RA-Attack), a novel jailbreak framework that fully exploits this vulnerability. RA-Attack employs a cross-modal safe anchor, whose core component is a structured visual mind map. This structured format provides the model with a pre-established, safety-biased reasoning chain that subtly induces it to rationalize and execute subsequent harmful intent. Extensive experiments across seven leading closed- and open-source MLRMs demonstrate the effectiveness of RA-Attack, achieving state-of-the-art jailbreak success rates--92% on Gemini-2.5-Pro and 82% on GPT-4o. Our findings reveal that cognitive biases can be systematically exploited to manipulate multimodal reasoning chains, establishing cognitive security as a critical and underexplored frontier in AI safety research. Warning: This paper contains unsafe examples.

Multi-modal Large Language Models (MLLMs) show promise in video understanding. However, their reasoning often suffers from thinking drift and weak temporal comprehension, even when enhanced by Reinforcement Learning (RL) techniques like Group Relative Policy Optimization (GRPO). Moreover, existing RL methods usually depend on Supervised Fine-Tuning (SFT), which requires costly Chain-of-Thought (CoT) annotation and multi-stage training, and enforces fixed reasoning paths, limiting MLLMs' ability to generalize and potentially inducing bias. To overcome these limitations, we introduce Summary-Driven Reinforcement Learning (SDRL), a novel single-stage RL framework that obviates the need for SFT by utilizing a Structured CoT format: Summarize - Think - Answer. SDRL introduces two self-supervised mechanisms integrated into the GRPO objective: 1) Consistency of Vision Knowledge (CVK) enforces factual grounding by reducing KL divergence among generated summaries; and 2) Dynamic Variety of Reasoning (DVR) promotes exploration by dynamically modulating thinking diversity based on group accuracy. This novel integration effectively balances alignment and exploration, supervising both the final answer and the reasoning process. Our method achieves state-of-the-art performance on seven public VideoQA datasets.

WiFi sensing offers passive and privacy-preserving perception that complements vision-based sensing, but its performance degrades sharply under domain shifts caused by changes in environment, subjects, or hardware. This challenge is exacerbated in real-world deployments where source data are unavailable, motivating test-time adaptation (TTA) as a practical solution for self-calibration using only unlabeled target samples. We introduce WiTTA-Bench, the first comprehensive benchmark for WiFi TTA, covering 20 representative methods, two adaptation protocols (OTTA and TTDA), and three major physics-induced shifts in WiFi: cross-environment, cross-subject, and cross-device. Furthermore, we contribute a new dataset featuring paired recordings from heterogeneous devices to bridge the cross-device gap. Extensive experiments reveal three key insights unique to WiFi sensing: (i) WiFi domain shifts exhibit a physics-induced hierarchy; (ii) OTTA and TTDA are complementary; (iii) OTTA is generally more robust to hyperparameters, while TTDA is more sensitive due to recursive self-training. WiTTA-Bench establishes the first systematic foundation for adaptive, robust, and deployable WiFi sensing under realistic wireless conditions.

In this paper, we tackle the Aerial Vision-and-Dialog Navigation (AVDN) task in the training-free setting for resource-efficient high-altitude UAV navigation.Naively applying MLLMs leads to unreliable navigation due to weak directional grounding and the lack of explicit spatial memory.To address these issues, we propose PSC-AVDN, a training-free framework that tightly couples a three-stage Parsing-Search-Confirmation reasoning pipeline with a Structured Spatial Memory (SSM).The parsing stage uses an LLM to convert ambiguous dialogue instructions into stable geometric directional and destination cues.A Search Chain-of-Thought (S-CoT) then performs stepwise target exploration under high-altitude observations, and a Confirmation Chain-of-Thought (C-CoT) conducts fine-grained verification around candidate regions to resolve visual ambiguity.Meanwhile, SSM integrates three complementary sources of spatial cues, including multi-scale visual observation, spatial visual memory, and structured geometric memory to provide global spatial context and long-horizon consistency.Extensive experiments on ANDH and ANDH-Full show that PSC-AVDN establishes new state-of-the-art performance in the training-free setting, matching or surpassing several finetuned methods.

Large Vision Language Models (LVLMs) exhibit strong perceptual and linguistic capabilities yet struggle with complex visual reasoning tasks that require structured, compositional, and adaptive inference. Existing approaches either rely on costly inference-time exploration--such as multi-path or tree-based Chain-of-Thought (CoT) search--or on expensive post-training with large curated CoT datasets. We propose ReaGEN, a lightweight framework for the adaptive generation of structured reasoning chains that enhances reasoning without modifying the underlying vision-language model (VLM). ReaGEN first employs a teacher-guided evolutionary search to collect sample specific CoT structure, leveraging attention-derived stage importance to capture how information flows across reasoning stages. These adaptive CoT structures are then used to train a compact generator (GEN) that learns to refine and improve CoT structures by reflecting on attention feedback from the reasoning process. At inference, the GEN dynamically produces question-adaptive structured CoTs, and can be iteratively invoked to refine them based on the VLM's internal state--achieving the flexibility of deep search with single-path efficiency. Across diverse multimodal reasoning benchmarks, ReaGEN achieves up to +26 accuracy points over test-time scaling methods while reducing the average inference-time token usage by 79%, making it a scalable approach for structured reasoning generation in self-hosted or open-source VLMs with access to internal attention signals.

Chain-of-Thought (CoT) is a key technique for enhancing the reasoning capabilities of Vision Language Models (VLMs). Existing methods often employ Reinforcement Learning (RL) with external constraints to align the model's reasoning process with human cognitive patterns. However, we argue that the model's intrinsic reasoning paths may differ from human cognition, and that forcing such alignment can constrain the model's potential and even degrade its performance. To address this, we propose leveraging the model's intrinsic self-evaluation to guide its optimization. We hypothesize that a model's self-generated confidence scores are effective indicators of its reasoning quality. Based on this evaluation metric, we design two novel reward functions: (1) Sequential Confidence Rigorous Evaluation (SCRE) for challenging problems that demand strict logical reasoning, and (2) Intra-group Score Re-ranking (IGSR) for general-purpose, open-ended scenarios. We name our method Video-RAISE (Reasoning Alignment through Intrinsic Self-Evaluation). Comprehensive experiments on six video understanding benchmarks demonstrate that Video-RAISE achieves state-of-the-art (SOTA) performance, significantly outperforming previous methods and even proprietary models, e,g. GPT-4o and Gemini 1.5 pro. For instance, on the VideoMMMU benchmark, our Video-RAISE achieves a new SOTA accuracy of 52.8%, outperforming the previous best model by a significant 3.0%. In addition, our method achieves a reasoning path consistency of 90%, which is double that of the Qwen2.5-VL-Instruct and even surpasses the performance of supervised fine-tuning.

While Large Multimodal Models (LMMs) have made significant progress, they remain largely text-centric, relying on language as their core reasoning modality. As a result, they are limited in their ability to handle reasoning tasks that are predominantly visual. Recent approaches have sought to address this by supervising intermediate visual steps with helper images, depth maps, or image crops. However, these strategies impose restrictive priors on what "useful" visual abstractions look like, add heavy annotation costs, and struggle to generalize across tasks. To address this critical limitation, we propose a task-agnostic mechanism that trains LMMs to discover and use visual reasoning tokens without explicit supervision. These tokens attend globally and re-encode the image in a task-adaptive way, enabling the model to extract relevant visual information without hand-crafted supervision. Our approach outperforms direct fine-tuning and achieves state-of-the-art results on a diverse range of vision-centric tasks -- including those where intermediate abstractions are hard to specify -- while also generalizing to multi-task instruction tuning.

The Abstraction and Reasoning Corpus (ARC) is designed to promote research on abstract reasoning, a fundamental aspect of human intelligence. Common approaches to ARC treat it as a language-oriented problem, addressed by large language models (LLMs) or recurrent reasoning models. However, although the puzzle-like tasks in ARC are inherently visual, existing research has rarely approached the problem from a vision-centric perspective. In this work, we formulate ARC within a vision paradigm, framing it as an image-to-image translation problem. To incorporate visual priors, we represent the inputs on a "canvas" that can be processed like natural images. It is then straightforward for us to apply standard vision architectures, such as a vanilla Vision Transformer (ViT), to perform image-to-image mapping. Our model is trained from scratch solely on ARC data and generalizes to unseen tasks through test-time training. Our framework, termed Vision ARC (VARC), achieves 60.4% accuracy on the ARC-1 benchmark, substantially outperforming existing methods that are also trained from scratch. Our results are competitive with those of leading LLMs and close the gap to average human performance.

Multimodal Emotion Analysis (MEA) is crucial for human-centric AI, yet current methods struggle with two core challenges: the sparse nature of emotional cues across modalities and their inherent temporal asynchrony. Existing approaches, which often rely on implicit fusion, consequently suffer from diluted salient features and entangled representations. To address this issue, we propose EmoThinker, a new framework that advances MEA through explicit, structured reasoning. Our method introduces a structural token selection mechanism to concentrate on pivotal facial regions while refining background context, enhancing visual saliency and efficiency. For audio, an audio evidence extractor aggregates critical paralinguistic features into compact, emotion-rich tokens. More importantly, we enable step-by-step reasoning by constructing a Chain-of-Emotion-Thought dataset, which provides fine-grained annotations for disentangling asynchronous cues and resolving inter-modal conflicts. By decoupling evidence acquisition from reasoning, EmoThinker achieves a more interpretable and robust emotion analysis. Extensive experiments on multiple benchmarks demonstrate that our framework achieves new state-of-the-art performance.

It is extremely hard for an underwater agent to know where it is. Satellite signals disappear within centimeters of the surface; acoustic baselines require heavy infrastructure to instrument small regions. The polarization of the sky, visible underwater, reveals the elevation of the sun. The pattern of elevation over the day reveals location to an agent with a clock. However, recovering elevation from polarization images is very difficult. State-of-the-art (SOTA) geolocalization methods can localize well for locations where they have seen data, but accuracy collapses when the data comes from a new location. Our physics-guided synthesis pipeline expands a huge library of polarization images from a small set of sites to 2.8 million solar-elevation-matched training sequences spanning latitudes, seasons, and water types. A compact two-stage transformer reconstructs the solar-elevation curve and predicts geolocation. Under leave-one-site-out tests, the site averaged median geodesic error is 500km--about an eightfold improvement over previous deep-learning baselines; with limited target-site data, the median error contracts to single-digit kilometers.