CVPR 2026 Papers

Unified interpreting and synthesizing human behaviors within 3D environments is vital for advancing spatial intelligence and humanoid robotics. Despite recent advancements (e.g., HSI-GPT), two fundamental capabilities expected of a unified model--understanding and generation--still lag behind specialist models. This is primarily due to 1) single-granularity codebook overemphasizes low-frequency motion details while neglecting motion semantics, 2) limited decoding capacity of the motion detokenizer which restricts the fidelity of human-scene interactions, 3) only relying on supervised fine-tuning (SFT) failing to capture high-level motion semantics and logical reasoning with an end-to-end mapping. To this issue, we develop HSI-GPT2--a reasoning-enhanced, dual-granularity motion-representational large Scene-Motion-Language model, powered by reinforcement learning (RL) with Chain-of-Thought (CoT) reasoning. First, HSI-GPT2 introduces a Dual-granularity Motion Tokenizer, DMoTok, which jointly preserves both fine-grained motion details and text-aligned motion semantics for various HSI-related tasks. Further, a motion diffusion decoder functions as a motion detokenizer, translating deep semantics and detailed features of LLMs into physically grounded human motions. Finally, we curate a Motion Chain-of-Thought (MoCoT) data engine and extend a Group Relative Policy Optimization (GRPO) paradigm to execute long-horizon and compositionallyrich commands. Results on standard HSI benchmarks confirm the clear superiority of HSI-GPT2 in enhancing interaction quality, semantic alignment, behavioral diversity, and generalization to unseen 3D scenes.

CoT has significantly enhanced the reasoning ability of LLMs while it faces challenges when extended to multimodal domains, particularly in mathematical tasks.Existing MLLMs typically perform textual reasoning solely from a single static mathematical image, overlooking dynamic visual acquisition during reasoning.In contrast, humans repeatedly examine visual image and employ step-by-step reasoning to prove intermediate propositions.This strategy of decomposing the problem-solving process into key logical nodes adheres to Miller's Law in cognitive science.Inspired by this insight, we propose a **ViRC** framework for multimodal mathematical tasks, introducing a **Reason Chunking** mechanism that structures multimodal mathematical CoT into consecutive **Critical Reasoning Units (CRUs)** to simulate human expert problem-solving patterns.CRUs ensure intra-unit textual coherence for intermediate proposition verification while integrating visual information across units to generate subsequent propositions and support structured reasoning.To this end, we present **CRUX** dataset by using three visual tools and four reasoning patterns to provide explicitly annotated CRUs across multiple reasoning paths for each mathematical problem.Leveraging the CRUX dataset, we propose a progressive training strategy inspired by human cognitive learning, which includes Instructional SFT, Practice SFT, and Strategic RL, aimed at further strengthening the Reason Chunking ability of the model.The resulting **ViRC-7B** model achieves a 18.8% average improvement over baselines across multiple mathematical benchmarks.The codes will be made publicly available.

Recent advancements in video models have shown tremendous progress, particularly in long video understanding. However, current benchmarks predominantly feature western-centric data and English as the dominant language, introducing significant biases in evaluation. To address this, we introduce CURVE, a challenging benchmark for multicultural and multilingual video reasoning. CURVE comprises high-quality, entirely human-generated annotations from diverse, region-specific cultural videos across 18 global locales. Unlike prior work that relies on automatic translations, CURVE provides complex questions, answers, and multi-step reasoning steps, all crafted in native languages. Making progress on CURVE requires a deeply situated understanding of visual cultural context. Furthermore, we leverage CURVE's reasoning traces to construct evidence-based graphs and propose a novel iterative strategy using these graphs to identify fine-grained errors in reasoning. Our evaluations reveal that SoTA Video-LLMs struggle significantly, performing substantially below human-level accuracy, with errors primarily stemming from the visual perception of cultural elements. We will release CURVE to foster the development of more equitable and capable multimodal foundation models.

We present Proof-of-Perception (PoP), a tool-using framework that casts multimodal reasoning as an executable graph with explicit reliability guarantees. Each perception or logic node outputs a conformal set \Gamma^ (t) _\delta(x), yielding calibrated, stepwise uncertainty; a lightweight controller uses these certificates to allocate compute under a budget--expanding with extra tool calls only when needed and stopping early otherwise. This grounds answers in verifiable evidence, reduces error compounding and hallucinations, and enables principled accuracy-compute trade-offs. Across document, chart, and multi-image QA benchmarks, PoP improves performance and reliability over strong chain-of-thought, ReAct-style, and program-of-thought baselines while using computation more efficiently. Code is available at \href https://github.com/AryaFayyazi/PoP https://github.com/AryaFayyazi/PoP .

Multimodal language models (MLLMs) are increasingly paired with vision tools (e.g., depth, flow, correspondence) to enhance visual reasoning. However, despite access to these tool-generated visual cues, MLLMs often fail to benefit fully from them. Existing approaches typically feed raw tool outputs into the model, but these dense, pixel-level representations are misaligned with the language-native reasoning strengths of LLMs, leading to weak perception due to reliance on language priors. We argue that, in problems where vision tools can provide the necessary visual cues, the bottleneck is not more tool calls or larger MLLMs, it is how tool outputs are represented. We introduce Perception Programs (P^2), a training-free, model-agnostic method that rewrites tool outputs into compact, structured, language-native summaries that MLLMs can directly parse and reason over. Across six perception-centric tasks in BLINK, P^2 consistently yields large improvements over base models and raw tool-augmented baselines. With GPT-5 Mini as the base model, P^2 raises its accuracy from 41.35% to 86.47% on multi-view reasoning, from 52.42 to 81.45 on relative depth, and achieves a 19.66 overall average gain across tasks, setting new state-of-the-art results. Even on smaller MLLMs, e.g., InternVL3.5-4B and Qwen3VL-4B, we observe 21-25 absolute gains from P^2 across tasks when comparing to image-based tool variants, surpassing prior agentic, supervised, and RL-based tool-use methods, without any training or model modifications.

Machine unlearning aims to erase requested data from trained models without full retraining. For Reasoning Multimodal Large Language Models (RMLLMs), this is uniquely challenging: intermediate chain-of-thought steps can still leak sensitive information even when final answers are forgotten, and overly aggressive interventions easily damage general reasoning ability. Yet no benchmark jointly evaluates how well unlearning methods suppress reasoning-level leakage while preserving reasoning competence. We address this gap with RMLLMU-Bench, the first benchmark for RMLLM unlearning that extends standard forgetting metrics with dedicated measures of reasoning leakage and reasoning retention. A systematic evaluation on RMLLMU-Bench reveals that existing unlearning methods for MLLMs and Large (Language) Reasoning Models (LRMs) either leave substantial leakage in the reasoning process or severely degrade reasoning performance. To address these gaps, we propose R-MUSE (\underline R easoning-preserving \underline M LLM \underline U nlearning via \underline S ubspace guidance and Adaptive St \underline E ering), a training-free and inference-time intervention framework that steers internal representations to forget both answers and reasoning traces while explicitly preserving general reasoning. Experiments on RMLLMU-Bench demonstrate that R-MUSE achieves a substantially better balance between effective forgetting and reasoning retention.

Online reconstruction of dynamic scenes aims to learn from streaming multi-view inputs under low-latency constraints. The fast training and real-time rendering capabilities of 3D Gaussian Splatting have made on-the-fly reconstruction practically feasible, enabling online 4D reconstruction. However, existing online approaches, despite their efficiency and visual quality, fail to learn per-Gaussian motion that reflects true scene dynamics. Without explicit motion cues, appearance and motion are optimized solely under photometric loss, causing per-Gaussian motion to chase pixel residuals rather than true 3D motion. To address this, we propose MoRGS, an efficient online per-Gaussian motion reasoning framework that treats Gaussian movement as a core modeling object. Specifically, we efficiently leverage optical flow on a sparse set of key views as a lightweight motion cue to guide per-Gaussian motion toward the scene's true dynamics. To compensate for the sparsity and view-dependence of flow, we learn a per-Gaussian motion offset field that reconciles discrepancies between projected 3D motion and observed flow across views and time. In addition, we introduce a per-Gaussian motion confidence that separates dynamic from static Gaussians and weights Gaussian attribute residual updates, thereby suppressing redundant motion in static regions for better temporal consistency and accelerating the modeling of large motions. Extensive experiments demonstrate that MoRGS achieves state-of-the-art reconstruction quality and motion fidelity among online methods, while maintaining streamable performance.

Reasoning segmentation enables open-set object segmentation via implicit text queries, therefore serving as a foundation for embodied agents that should operate autonomously in real-world environments. However, existing methods for reasoning segmentation require multimodal large language models with billions of parameters that exceed the computational capabilities of edge devices that typically deploy the embodied AI systems. Distillation offers a pathway to compress these models while preserving their capabilities. Yet, existing distillation approaches fail to transfer the multi-step reasoning capabilities that reasoning segmentation demands, as they focus on matching output predictions and intermediate features rather than preserving reasoning chains. The emerging paradigm of reasoning over digital twin representations presents an opportunity for more effective distillation by re-framing the problem. Consequently, we propose FastReasonSeg, which employs digital twin representations that decouple perception from reasoning to enable more effective distillation. Our distillation scheme first relies on supervised fine-tuning on teacher-generated reasoning chains. Then it is followed by reinforcement fine-tuning with joint rewards evaluating both segmentation accuracy and reasoning quality alignment. Experiments on two video (JiTBench, RVTBench) and two image benchmarks (ReasonSeg, LLM-Seg40K) demonstrate that our FastReasonSeg achieves state-of-the-art reasoning segmentation performance. Moreover, the distilled 0.6B variant outperforms models with 20 times more parameters while achieving 7.79 FPS throughput with only 2.1GB memory consumption. This efficiency enables deployment in resource-constrained environments to enable real-time reasoning segmentation.

Face recognition remains vulnerable to presentation attacks, calling for robust Face Anti-Spoofing (FAS) solutions. Recent MLLM-based FAS methods reformulate the binary classification task as the generation of brief textual descriptions to improve cross-domain generalization. However, their generalizability is still limited, as such descriptions mainly capture intuitive semantic cues (e.g., mask contours) while struggling to perceive fine-grained visual patterns. To address this limitation, we incorporate external visual tools into MLLMs to encourage deeper investigation of subtle spoof clues. Specifically, we propose the Tool-Augmented Reasoning FAS (TAR-FAS) framework, which reformulates the FAS task as a Chain-of-Thought with Visual Tools (CoT-VT) paradigm, allowing MLLMs to begin with intuitive observations and adaptively invoke external visual tools for fine-grained investigation. To this end, we design a tool-augmented data annotation pipeline and construct the ToolFAS-16K dataset, which contains multi-turn tool-use reasoning trajectories. Furthermore, we introduce a tool-aware FAS training pipeline, where Diverse-Tool Group Relative Policy Optimization (DT-GRPO) enables the model to autonomously learn efficient tool use. Extensive experiments under a challenging one-to-eleven cross-domain protocol demonstrate that TAR-FAS achieves SOTA performance while providing fine-grained visual investigation for trustworthy spoof detection.

Large vision-language models (VLMs) show strong multimodal understanding but still struggle with 3D spatial reasoning, such as distance estimation, size comparison, and cross-view consistency. Existing 3D-aware methods either depend on auxiliary 3D information or enhance RGB-only VLMs with geometry encoders through shallow feature fusion. We propose SpaceMind, a multimodal large language model trained on SpaceMind-700K for spatial reasoning from RGB inputs. The model adopts a dual-encoder architecture integrating VGGT as a spatial understanding encoder and InternViT as a 2D visual encoder. The key idea is to treat the camera representation as an active guiding modality rather than passive metadata. Specifically, SpaceMind introduces a lightweight Camera-Guided Modality Fusion module before the language model to replace shallow fusion. It applies camera-conditioned biasing to spatial tokens, assigns query-independent weights reflecting their geometric importance, and uses the camera embedding to gate the fused representation. Empirically, SpaceMind establishes new state-of-the-art results on VSI-Bench, SQA3D and SPBench, surpassing both open and proprietary systems on VSI-Bench and SPBench by large margins and achieving state-of-the-art performance on SQA3D. These results demonstrate that camera-guided modality fusion is an effective and practical inductive bias for equipping VLMs with genuinely spatially grounded intelligence. We will release code and model checkpoints to support future research.

Chain-of-thought (CoT) reasoning has emerged as a powerful tool for multimodal large language models on video understanding tasks. However, its necessity and advantages over direct answering remain underexplored. In this paper, we first demonstrate that for RL-trained video models, direct answering often matches or even surpasses CoT performance, despite CoT producing step-by-step analyses at a higher computational cost. Motivated by this, we propose VideoAuto-R1, a video understanding framework that adopts a reason-when-necessary strategy. During training, our approach follows a Thinking Once, Answering Twice paradigm: the model first generates an initial answer, then performs reasoning, and finally outputs a reviewed answer. Both answers are supervised via verifiable rewards. During inference, the model uses the confidence score of the initial answer to determine whether to proceed with reasoning. Across video QA and grounding benchmarks, VideoAuto-R1 achieves state-of-the-art accuracy with significantly improved efficiency, reducing the average response length by 3.3x, eg., from 149 to just 44 tokens. Moreover, we observe a low rate of thinking-mode activation on perception-oriented tasks, but a higher rate on reasoning-intensive tasks. This suggests that explicit language-based reasoning is generally beneficial but not always necessary.

In this paper, we present a holistic multimodal benchmark that evaluates the reasoning capabilities of MLLMs with an explicit focus on reasoning width, a complementary dimension to the more commonly studied reasoning depth. Specifically, reasoning depth measures the model's ability to carry out long-chain, sequential reasoning in which each step is tightly and rigorously linked to the next. Reasoning width tends to focus more on the model's capacity for broad trial-and-error search or multi-constrained optimization: it must systematically traverse many possible and parallelized reasoning paths, apply diverse constraints to prune unpromising branches, and identify valid solution routes for efficient iteration or backtracking. To achieve it, we carefully curate 1200+ high-quality multimodal cases spanning heterogeneous domains, and propose a fine-grained tree-of-thought evaluation protocol that jointly quantifies reasoning width and depth. We evaluate 12 major model families (over 30 advanced MLLMs) across difficulty tiers, question types, and required skills. Results show that while current models exhibit strong performance on general or common-sense VQA tasks, they still struggle to combine deep sequential thought chains with wide exploratory search to perform genuine insight-based reasoning. Finally, we analyze characteristic failure modes to provide possible directions for building MLLMs that reason not only deeper but also wider.

Abstract reasoning from minimal examples remains a core unsolved problem for frontier foundation models such as GPT-5. These models still fail to infer structured transformation rules from a handful of examples, which is a key hallmark of human intelligence. The Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI) provides a rigorous testbed for this capability, demanding conceptual rule induction and transfer to novel tasks. Most existing methods treat ARC-AGI as a purely textual reasoning task, overlooking the fact that humans rely heavily on visual abstraction when solving such puzzles. However, our experiments reveal a paradox: naively rendering ARC-AGI grids as images degrades performance due to imprecise rule execution. This leads to our central hypothesis that vision and language possess complementary strengths across distinct reasoning stages: vision supports global pattern abstraction and verification, whereas language specializes in symbolic rule formulation and precise execution. Building on this insight, we introduce two synergistic strategies: (1) Vision-Language Synergy Reasoning (VLSR), which decomposes ARC-AGI into modality-aligned subtasks; and (2) Modality-Switch Self-Correction (MSSC), which leverages vision to verify text-based reasoning for intrinsic error correction. Extensive experiments demonstrate that our approach yields up to a 4.33% improvement over text-only baselines across diverse flagship models and multiple ARC-AGI tasks. Our findings suggest that unifying visual abstraction with linguistic reasoning is a crucial step toward achieving generalizable, human-like intelligence.

We propose MIRA (Multimodal Imagination for Reasoning Assessment), a new benchmark designed to evaluate models in scenarios where generating intermediate visual images is essential for successful reasoning. Unlike traditional Chain-of-thought (CoT) methods that rely solely on text, tasks in MIRA require models to generate and utilize intermediate images --- such as sketches, structural diagrams, or path drawings --- to guide their reasoning process. This setup closely mirrors how humans solve complex problems through "drawing to think". To solve this, MIRA focuses on tasks that are intrinsically challenging and involve complex structures, spatial relationships, or reasoning steps that are difficult to express through language alone (e.g., tracking a die's movement on a board and summing the face-down values after each roll). To ensure that our evaluation data is of high-quality, we include 546 multimodal problems, annotated with intermediate visual images and final answers. We also propose a unified evaluation protocol for MIRA that spans three levels of evaluation input: direct input with image and question only, text-only CoT (Text-CoT) input with image and thinking prompts, and Visual-CoT input with both annotated image clues and textual thinking prompts. To probe the upper bound of model capacity on our benchmark, we also report pass@k and majority voting accuracies under different k settings. Experimental results show that existing multimodal large language models (MLLMs), including strongest private models (e.g., GPT-5, o3, Gemini 2.5 Pro) as well as strong open-weight models (e.g., Qwen2.5-VL, GLM 4.5V), perform poorly when relying solely on textual prompts. However, when intermediate visual cues are provided, model performance improves consistently, yielding an average relative gain of 33.7% across all models and tasks. We also probe the upper bound by expanding the search space and designing textual prompts aligned with Visual-CoT, but both yield only limited improvements compared to our Visual-CoT setting. These results underscore the critical role of imagined visual information in enabling successful reasoning on MIRA.

Self-reflection mechanisms that rely on purely text-based rethinking processes perform well in most multimodal tasks. However, when directly applied to long-form video understanding scenarios, they exhibit clear limitations. The fundamental reasons for this lie in two points: (1) long-form video understanding involves richer and more dynamic visual input, meaning rethinking only the text information is insufficient and necessitates a further rethinking process specifically targeting visual information; (2) purely text-based reflection mechanisms lack cross-modal interaction capabilities, preventing them from fully integrating visual information during reflection. Motivated by these insights, we propose REVISOR (REflective VIsual Segment Oriented Reasoning), a novel framework for tool-augmented multimodal reflection. REVISOR enables MLLMs to collaboratively construct introspective reflection processes across textual and visual modalities, significantly enhancing their reasoning capability for long-form video understanding. To ensure that REVISOR can learn to accurately review video segments highly relevant to the question during reinforcement learning, we designed the Dual Attribution Decoupled Reward (DADR) mechanism. Integrated into the GRPO training strategy, this mechanism enforces causal alignment between the model's reasoning and the selected video evidence. Notably, the REVISOR framework significantly enhances long-form video understanding capability of MLLMs without requiring supplementary supervised fine-tuning or external models, achieving impressive results on four benchmarks including VideoMME, LongVideoBench, MLVU, and LVBench.

Abstract visual reasoning benchmarks such as ARC-AGI evaluate the ability to infer generalizable transformation rules from few graphical demonstrations, a capability where current deep learning models severely underperform. Mainstream LLMs achieve only 15.8% (DeepSeek-R1) and 34.5% (o3-mini-high) accuracy. The core reason lies in their static processing of task examples: unlike humans, who iteratively refine their understanding of examples while solving problems, these models lack mechanisms for dynamically aligning understanding and solving. We address this gap with the Understanding and Solving Reasoning Loop (USRL) framework. The architecture comprises two explicitly interacting modules: an Understanding Module (UM) that encodes and refines rule representations of examples, and a Solving Module (SM) that generates a draft solution informed by these evolving rule representations. Through recurrent interaction, the model iteratively aligns its draft solution with its understanding about task examples continuously. Furthermore, we introduce an adaptive reasoning halting mechanism that autonomously terminates the reasoning loop based on the consistency between the generated draft solution and the examples. With 7M parameters, our model achieves 47.2% accuracy on ARC-AGI-1.

Many audio-visual learning methods have focused on aligning audio and visual information, either through semantic or temporal correspondence. However, most of these works have utilized monaural audio, which does not contain information about the spatial location of the sound source. In contrast, humans and other animals utilize binaural hearing to perceive this spatial information. Combining spatial sound and visual perception enables powerful high-level reasoning: for example, a person looking for their phone may hear the ringing sound coming from a backpack sitting on a table, and quickly infer that the missing phone is inside the backpack. In this paper, we investigate the problem of Audio-Visual Spatial Reasoning. We design a spatial audio-visual question answering dataset to cover scenarios where semantic correspondence between audio and visual signals is absent but spatial alignment exists, as well as cases with multiple audio-visual semantic correspondences that require spatial reasoning to disambiguate. We propose a model that learns spatial comprehension across the audio and vision modalities by connecting them with a large language model and experimentally demonstrate that spatial sound perception is an essential part of our task.

Recent advances in 3D scene-language understanding have leveraged Large Language Models (LLMs) for 3D reasoning by transferring their general reasoning ability to 3D multi-modal contexts. However, existing methods typically adopt standard decoders from language modeling, which rely on a causal attention mask. This design introduces two fundamental conflicts in 3D scene understanding: sequential bias among order-agnostic 3D objects and restricted object-instruction attention, hindering task-specific reasoning. To overcome these limitations, we propose 3D Spatial Language Instruction Mask (3D-SLIM), an effective masking strategy that replaces the causal mask with an adaptive attention mask tailored to the spatial structure of 3D scenes. Our 3D-SLIM introduces two key components: a Geometry-adaptive Mask that constrains attention based on spatial density rather than token order, and an Instruction-aware Mask that enables object tokens to directly access instruction context. This design allows the model to process objects based on their spatial relationships while being guided by the user's task. 3D-SLIM is simple, requires no architectural modifications, and adds no extra parameters, yet it yields substantial performance improvements across diverse 3D scene-language tasks. Extensive experiments across multiple benchmarks and LLM baselines validate its effectiveness and underscore the critical role of decoder design in 3D multi-modal reasoning.

Large vision-language models (VLMs) often benefit from intermediate visual cues, either injected via external tools or generated as latent visual tokens during reasoning, but these mechanisms still overlook fine-grained visual evidence (e.g., polylines in charts), generalize poorly across domains, and incur high inference-time cost. In this paper, we propose Bi-directional Perceptual Shaping (BiPS), which transforms question-conditioned masked views into bidirectional where-to-look signals that shape perception during training. BiPS first applies a KL-consistency constraint between the original image and an evidence-preserving view that keeps only question-relevant regions, encouraging coarse but complete coverage of supporting pixels. It then applies a KL-separation constraint between the original and an evidence-ablated view where critical pixels are masked so the image no longer supports the original answer, discouraging text-only shortcuts (i.e., answering from text alone) and enforcing fine-grained visual reliance. Across eight benchmarks, BiPS boosts Qwen2.5-VL-7B by 8.2% on average and shows strong out-of-domain generalization to unseen datasets and image types.