CVPR 2026 Papers

Multimodal Large Language Models (MLLMs) with explicit step-by-step reasoning have achieved strong performance on complex tasks. However, such reasoning is unnecessary for many simple queries and introduces substantial computational overhead. To address this inefficiency, we present R-4B, an auto-thinking MLLM that dynamically determines whether to invoke the reasoning process based on input complexity.Our key idea is to equip a single model with both thinking and non-thinking capabilities and train it to select the appropriate mode. We first introduce bi-mode annealing, a unified training paradigm that constructs a model competent in both reasoning-intensive and direct-answer settings without requiring explicit complexity annotations. Building on this foundation, we propose Bi-mode Policy Optimization (BPO), a lightweight reinforcement learning algorithm that employs a dual-rollout mechanism: for each input, the model generates both thinking and non-thinking responses. This prevents mode collapse and enables robust learning of an adaptive reasoning policy using only simple, rule-based rewards.Extensive experiments across 25 benchmarks show that R-4B achieves state-of-the-art performance among models of similar scale. It consistently surpasses Qwen2.5-VL-7B and matches or exceeds larger models such as Kimi-VL-A3B-Thinking-2506 (16B) on reasoning-intensive tasks, while reducing computational cost by avoiding redundant reasoning. Our results demonstrate that adaptive auto-thinking offers an effective and scalable pathway toward more efficient multimodal reasoning models.

Text-goal instance navigation (TGIN) asks an agent to resolve a single, free-form description into actions that reach the correct object instance among same-category distractors. We present Context-Nav, which elevates long, contextual captions from a local matching cue to a global exploration prior and verifies candidates through 3D spatial reasoning. First, we compute dense text-image alignments for a value map that ranks frontiers---guiding exploration toward regions consistent with the entire description rather than early detections. Second, upon observing a candidate, we perform a viewpoint-aware relation check: the agent samples plausible observer poses, aligns local frames, and accepts a target only if the spatial relations can be satisfied from at least one viewpoint. The pipeline requires no task-specific training or fine-tuning; we attain state-of-the-art performance on InstanceNav and CoIN-Bench. Ablations show that (i) encoding full captions into the value map avoids wasted motion and (ii) explicit, viewpoint-aware 3D verification prevents semantically plausible but incorrect stops. This suggests that geometry-grounded spatial reasoning is a scalable alternative to heavy policy training or human-in-the-loop interaction for fine-grained instance disambiguation in cluttered 3D scenes.

Recent releases such as o3 highlight human-like "thinking with images" reasoning that combines tool use with stepwise verification, yet most open-source approaches still rely on text-only chains, rigid visual schemas, or single-step pipelines, limiting flexibility, interpretability, and transferability on complex tasks. We introduce CodeDance, which explores executable code as a general solver for visual reasoning. Unlike fixed-schema calls (e.g., only predicting bounding-box coordinates), CodeDance defines, composes, and executes code to orchestrate multiple tools, compute intermediate results, and render visual artifacts (e.g., boxes, lines, plots) that support transparent, self-checkable reasoning. To guide this process, we introduce a reward for balanced and adaptive tool calling, which balances exploration with efficiency and mitigates tool overuse. Interestingly, beyond the expected capabilities taught by atomic supervision, we empirically observe novel emergent behaviors during RL training: CodeDance demonstrates novel tool invocations, unseen compositions, and cross-task transfer. These behaviors arise without task-specific fine-tuning, suggesting a general and scalable mechanism for executable visual reasoning. Extensive experiments across reasoning benchmarks (e.g., visual search, math, chart QA) show that CodeDance not only consistently outperforms schema-driven and text-only baselines, but also surpasses closed models such as GPT-4o and larger open-source models.

Long video understanding is challenging due to dense visual redundancy, long-range temporal dependencies, and the tendency of chain-of-thought and retrieval-based agents to accumulate semantic drift and correlation-driven errors. We argue that long-video reasoning should begin not with reactive retrieval, but with deliberate task formulation: the model must first articulate what must be true in the video for each candidate answer to hold. This thinking-before-finding principle motivates VideoHV-Agent, a framework that reformulates video question answering as a structured hypothesis-verification process.Based on video summaries, a Thinker rewrites answer candidates into testable hypotheses, a Judge derives a discriminative clue specifying what evidence must be checked, a Verifier grounds and tests the clue using localized, fine-grained video content, and an Answer agent integrates validated evidence to produce the final answer.Experiments on three long-video understanding benchmarks show that VideoHV-Agent achieves state-of-the-art accuracy while providing enhanced interpretability, improved logical soundness, and lower computational cost.

Continual Test-Time Adaptation (CTTA) aims to empower perception systems to handle dynamic distribution shifts encountered after deployment. Existing methods predominantly follow a backward-alignment paradigm, which rigidly aligns incoming data with supervisory surrogates derived from the source domain. Consequently, they struggle with unreliable supervision and evolving distribution shifts. To overcome these limitations, we introduce a novel forward-facilitation paradigm through a method termed Dynamic Style Bridging. Prior to deployment, we construct a compact knowledge base of generated class exemplars. During test time, to mitigate inherent generative bias and adapt these proxies to incoming data, we propose a multi-level bridging mechanism. This mechanism dynamically injects the proxies with incoming data styles at the input, statistical, and representation levels, while preserving the original semantics of the proxies. These high-fidelity proxies are then used to provide reliable, on-demand supervisory signals, enabling stable adaptation under continual shifts. Extensive experiments across standard CTTA benchmarks demonstrate that our method achieves consistent and substantial improvements over recent state-of-the-art approaches. Code is available at https://github.com/z1358/DAS.

Test-time adaptation (TTA) aims to mitigate performance degradation under distribution shifts by updating model parameters during inference. Existing approaches have primarily framed adaptation around affine modulation, focusing on recalibrating normalization layers. This perspective, while effective, overlooks another influential component in representation dynamics: the activation function. We revisit this overlooked space and propose AcTTA, an activation-aware framework that reinterprets conventional activation functions from a learnable perspective and updates them adaptively at test time. AcTTA reformulates conventional activation functions (e.g., ReLU, GELU) into parameterized forms that shift their response threshold and modulate gradient sensitivity, enabling the network to adjust activation behavior under domain shifts. This functional reparameterization enables continuous adjustment of activation behavior without modifying network weights or requiring source data. Despite its simplicity, AcTTA achieves robust and stable adaptation across diverse corruptions. Across CIFAR10-C, CIFAR100-C, and ImageNet-C, AcTTA consistently surpasses normalization-based TTA methods. Our findings highlight activation adaptation as a compact and effective route toward domain-shift-robust test-time learning, broadening the prevailing affine-centric view of adaptation.

Text-image to video generation aims to synthesize a video conditioned on the given text-image inputs. Nevertheless, existing methods generally assume that the semantic information carried in the input text and image tends to be perfectly paired and temporally aligned, occurring simultaneously in the generated video. As such, existing literature struggles with out-of-distribution (OOD) "unpaired" text-image inputs in the more universal and realistic scenario where i) the semantic information carried by the text and image may occur at different timestamps and ii) the condition image can appear at an arbitrary position rather than the first frame of the synthesized video. Video generation under this OOD setting poses an urgent need to conduct reasoning over the intrinsic connections between the given textual description and referred image, which is challenging and remains unexplored. To address the challenge, in this paper we study the problem of unpaired text-image to video generation for the first time, proposing ReasonDiff, a novel model for accurate video generation from unpaired text-image inputs. Specifically, ReasonDiff designs a VisionNarrator module to harness the powerful reasoning abilities of a multi-modal LLM to analyze the unpaired text-image inputs, producing coherent per-frame narratives that temporally align them. Building upon this VisionNarrator module, ReasonDiff further introduces a novel AlignFormer module, which employs a Multi-stage Temporal Anchor Attention mechanism to predict frame-wise latent representations. These reasoning-enhanced latents are subsequently fused with the condition frame, providing structured guidance throughout the video generation process. Extensive experiments and ablation studies demonstrate that ReasonDiff beats state-of-the-art baselines in terms of video generation quality with unpaired text-image inputs.

Vision-Language-Action (VLA) models have recently emerged, demonstrating strong generalization in robotic scene understanding and manipulation. However, when confronted with long-horizon tasks that require defined goal states, such as LEGO assembly or object rearrangement, existing VLA models still face challenges in coordinating long-horizon planning with precise manipulation.Therefore, we aim to endow a VLA model with the capability to infer the "how" process from the "what" outcomes, transforming goal states into executable procedures. In this paper, we introduce ManualVLA, a unified VLA framework built upon a Mixture-of-Transformers (MoT) architecture, enabling coherent collaboration between multimodal manual generation and action execution. Unlike prior VLA models that directly map sensory inputs to actions, we first equip ManualVLA with a planning expert that generates intermediate manuals consisting of images, visual prompts, and textual instructions. Building upon these multimodal manuals, we design a Manual Chain-of-Thought (ManualCoT) reasoning process that feeds them into the action expert, where each manual step provides explicit control conditions, while its latent representation offers implicit guidance for accurate manipulation. To alleviate the burden of data collection, we develop a high-fidelity digital-twin toolkit based on 3D Gaussian Splatting, which automatically generates manual data for planning expert training. ManualVLA demonstrates strong real-world performance, achieving an average success rate 32% higher than the previous hierarchical SOTA baseline on LEGO assembly and object rearrangement tasks.

Charts are high-density visual carriers of complex data and medium for information extraction and analysis. Due to the need for precise and complex visual reasoning, automated chart understanding poses a significant challenge to existing Multimodal Large Language Models (MLLMs). Many MLLMs trained with reinforcement learning (RL) face the challenge of credit assignment. Their advantage estimation, typically performed at the trajectory level, cannot distinguish between correct and incorrect reasoning steps within a single generated response. To address this limitation, we introduce SketchVL, a novel MLLM that optimized with FinePO, a new RL algorithm designed for fine-grained credit assignment within each trajectory. SketchVL's methodology involves drawing its intermediate reasoning steps as markers on the image and feeding the annotated image back to itself, creating a robust, multi-step reasoning process. During training, the FinePO algorithm leverages a Fine-grained Process Reward Model (FinePRM) to score each drawing action within a trajectory, thereby precisely assigning credit for each step. This mechanism allows FinePO to more strongly reward correct tokens when a trajectory is globally successful, and more heavily penalize incorrect tokens when the trajectory is globally suboptimal, thus achieving fine-grained reinforcement signals. Experiments show that SketchVL learns to align its step-level behavior with the FinePRM, achieving an average performance gain of 7.23% over its base model across chart datasets, natural image datasets, and mathematics, providing a promising new direction for training powerful reasoning models.

Video-based spatial reasoning -- such as estimating distances, judging directions, or understanding layouts from multiple views -- requires selecting informative frames and, when needed, actively seeking additional viewpoints during inference. Existing multimodal large language models (MLLMs) consume a fixed set of uniformly sampled frames and cannot request new views once reasoning begins, often missing the geometric cues necessary for reliable spatial judgments. We present EagleVision, a dual-stage framework that combines geometry-aware frame selection with active, Bird's-Eye-View (BEV)-grounded reasoning. In the first stage (macro perception), a semantics-perspective-fusion determinantal point process (SPF-DPP) selects a compact set of keyframes that jointly maximize semantic relevance and viewpoint diversity under a fixed token budget. In the second stage (micro verification), the model performs iterative spatial Chain-of-Thought: at each step it can either reason in text or predict a pose on the BEV plane to retrieve the nearest real frame, forming a closed-loop hypothesize-look-verify cycle. The querying policy is trained purely via reinforcement learning with a spatial grounding reward, requiring no human-annotated reasoning traces. On VSI-Bench and SQA3D, EagleVision achieves state-of-the-art performance among open-source vision-language models.

Long-wave infrared radiation captured by a thermal camera includes (a) emission from an object governed by its temperature and emissivity, and (b) reflected radiation from the surrounding environment. Separating these components is a long-standing challenge in thermography. Even when using multiple bands, the problem is under-determined without priors on emissivity. This difficulty is amplified in near ambient conditions, where emitted and reflected signals are of comparable magnitude. We present a dual-band video thermography framework that reduces this ambiguity by combining two complementary ideas at a per-pixel level: (i) spectral cues (ratio of emissivity between bands is unknown but fixed), and (ii) temporal cues (object radiation changes smoothly while background radiation changes rapidly). We derive an image formation model and an algorithm to jointly estimate the object's emissivity at each band, and the time-varying object and background temperatures. Experiments with calibrated and uncalibrated emissivities in everyday scenes (e.g., coffee pot heating up, palm print on mirrors, reflections of moving people), demonstrate robust separation and recovery of temperature fields.

Chart Question Answering (CQA) benchmarks are critical for evaluating Multimodal Large Language Models (MLLMs) on visual data reasoning. Existing benchmarks focus mainly on final-answer correctness, ignoring intermediate reasoning steps and the propagation of errors in multi-step processes. To address this, we introduce ChartR, a benchmark designed to assess both the accuracy and robustness of reasoning in chart-understanding tasks. Each question is decomposed into 4-10 sub-questions covering key reasoning types, and each chart includes four visually perturbed variants (blurred, noise-added, watermark-added, annotation-removed) to systematically evaluate robustness. ChartR contains 200 base charts, 800 variants, 1,652 questions, and 8,260 image-question pairs. We further propose a comprehensive evaluation framework with eight metrics that evaluate reasoning-chain accuracy, robustness under visual perturbations, and enable analysis of potential error propagation patterns. Experiments on twelve MLLMs, including general-purpose and chart-specialized models, reveal low reasoning reliability, early-step errors that may propagate, value extraction as the primary bottleneck, and sharp performance drops under perturbations, highlighting reliance on textual cues over true visual understanding.

Large Vision Language Models (LVLMs) have demonstrated remarkable capabilities in Chain-of-Thought (CoT) reasoning. However, existing LVLM reasoning paradigms only begin reasoning after the entire video becomes available, introducing unnecessary latency and diminishing attention to early visual cues in dynamic scenes. Inspired by the human ability to think while watching, we introduce a streaming reasoning paradigm for LVLMs, where reasoning unfolds sequentially with incoming frames and deepens after the full video is observed. We instantiate this paradigm through Think-as-You-See (TaYS), a unified framework that enables LVLMs to reason while watching by integrating streaming CoT generation, stream-constrained training, and stream-parallel inference. Specifically, TaYS employs temporally aligned streaming reasoning units with precise CoT supervision, enforces ordered reasoning via streaming attention masks and positional encodings, and utilizes a parallel KV caches mechanism that decouples input encoding from reasoning generation, ensuring alignment and true concurrency. We evaluate TaYS on the Qwen2.5-VL model family across representative video CoT tasks, including event dynamics analysis, causal reasoning, and thematic understanding. Experimental results show that TaYS achieves superior reasoning performance compared with batch-mode CoT, while reducing pre-reasoning latency to under one second and overall answer delay by more than 50%. These findings demonstrate the effectiveness of the streaming paradigm in enabling real-time, human-like reasoning for LVLMs.

Video reasoning models are a core component of egocentric and embodied agents. However, standard benchmarks for assessing models provide only evaluation of the output (e.g. the answer to a question), without evaluation of inter- mediate reasoning steps, and most provide answers only in the text domain. We introduce Minerva-Ego, a bench- mark for evaluating complex egocentric visual reasoning. We extend recent high-quality video data sources recorded from egocentric / embodied settings with a set of challenging, multi-step multimodal questions and spatiotemporally-dense human-annotated reasoning traces. Benchmarking experiments show that state-of-the-art models still have a large gap to human performance. To investigate this gap in detail, we annotate each reasoning trace in the dataset with the objects of interest required to solve the question, as spatio-temporal mask annotations. Through extensive evaluations, we identify that prompting frontier models with hints of 'where' and when to look yields substantial improvements in performance.

Unified multi-modal understanding/generative models have shown improved image editing performance by incorporating fine-grained understanding into their Chain-of-Thought (CoT) process. However, a critical question remains underexplored: what forms of CoT and training strategy can jointly enhance both the understanding granularity and generalization? To address this, we propose Meta-CoT, a paradigm that performs a two-level decomposition of any single-image editing operation with two key properties: (1) Decomposability. We observe that any editing intention can be represented as a triplet -- (task, target, required understanding ability). Inspired by this, Meta-CoT decomposes both the editing task and the target, generating task-specific CoT and traversing editing operations on all targets. This decomposition enhances the model's understanding granularity of editing operations and guides it to learn each element of the triplet during training, substantially improving the editing capability.(2) Generalizability. In the second decomposition level, we further break down editing tasks into five fundamental meta-tasks. We find that training on these five meta-tasks, together with the other two elements of the triplet, is sufficient to achieve strong generalization across diverse, unseen editing tasks.To further align the model's editing behavior with its CoT reasoning, we introduce the CoT-Editing Consistency Reward, which encourages more accurate and effective utilization of CoT information during editing. Experiments demonstrate that our method achieves an overall 15.8% improvement across 21 editing tasks, and generalizes effectively to unseen editing tasks when trained on only a small set of meta-tasks. Our code, benchmark, and model will be released publicly.

The rapid evolution of video generative models has shifted their focus from producing visually plausible outputs to tackling tasks requiring physical plausibility and logical consistency. However, despite recent breakthroughs such as Veo 3's chain-of-frames reasoning, it remains unclear whether these models can exhibit reasoning capabilities similar to large language models (LLMs). Existing benchmarks predominantly evaluate visual fidelity and temporal coherence, failing to capture higher-order reasoning abilities. To bridge this gap, we propose **TiViBench**, a hierarchical benchmark specifically designed to evaluate the reasoning capabilities of image-to-video (I2V) generation models. TiViBench systematically assesses reasoning across four dimensions: i) **Structural Reasoning & Search**, ii) **Spatial & Visual Pattern Reasoning**, iii) **Symbolic & Logical Reasoning**, and iv) **Action Planning & Task Execution**, spanning 24 diverse task scenarios across 3 difficulty levels. Through extensive evaluations, we show that commercial models (*e.g.*, Sora 2, Veo 3.1) demonstrate stronger reasoning potential, while open-source models reveal untapped potential that remains hindered by limited training scale and data diversity. To further unlock this potential, we introduce **VideoTPO**, a simple yet effective test-time strategy inspired by preference optimization. By performing LLM self-analysis on generated candidates to identify strengths and weaknesses, VideoTPO significantly enhances reasoning performance without requiring additional training, data, or reward models. Together, TiViBench and VideoTPO pave the way for evaluating and advancing reasoning in video generation models, setting a foundation for future research in this emerging field.

Despite the rapid progress in multimodal models and Large Visual-Language Models (LVLM), they remain highly susceptible to adversarial perturbations, raising serious concerns about their reliability in real-world use. While adversarial training has become the leading paradigm for building models that are robust to adversarial attacks, Test-Time Transformations (TTT) have emerged as a promising strategy to boost robustness at inference.In light of this, we propose Energy-Guided Test-Time Transformation (ET3), a lightweight, training-free defense that enhances the robustness by minimizing the energy of the input samples.Our method is grounded in a theory that proves our transformation succeeds in classification under reasonable assumptions. We present extensive experiments demonstrating that ET3 provides a strong defense for classifiers, zero-shot classification with CLIP, and also for boosting the robustness of LVLMs in tasks such as Image Captioning and Visual Question Answering. Code is available at github.com/OmnAI-Lab/Energy-Guided-Test-Time-Defense .

Streaming video reasoning requires models to operate in a setting where history grows without bound while meaningful evidence remains scarce. In such a landscape, relevant signal is like an oasis -- small, critical, and easily lost in a desert of redundancy. Enlarging memory only widens the desert; aggressive compression dries up the oasis. The real difficulty lies in discovering where to look, not how much to remember. We therefore introduce OASIS, a novel framework for streaming video reasoning that tackles this challenge through structured, on-demand retrieval. It organizes streaming history into hierarchical events and performs reasoning as controlled refinement -- short-context inference first, followed by semantically grounded retrieval only when uncertainty arises. As the retrieval is driven by high-level intent rather than embedding similarity, the retrieved memory is substantially more accurate and less noisy. Additionally, the mechanism is plug-and-play, training-free, and readily attaches to different streaming MLLM backbones. Experiments across multiple benchmarks and backbones show that OASIS achieves strong gains in long-horizon accuracy and compositional reasoning with bounded token cost and low request delay.

Humans inhabit a physical 4D world, where spatial geometry and semantic content evolve over time, forming a dynamic reality. While current Multimodal Large Language Models (MLLMs) demonstrate strong capabilities in understanding static visual inputs, it remains unclear whether they can effectively "think in dynamics," i.e., perceive, track, and reason about spatio-temporal evolution in complex scenes.To systematically evaluate these abilities, we introduce \texttt Dyn-Bench , a large-scale benchmark designed to assess spatio-temporal reasoning and localized dynamics perception. Constructed through multi-stage filtering over massive 2D and 4D data sources, \texttt Dyn-Bench provides a high-quality collection of diverse dynamic scenes, consisting of 1k videos, 7k visual question answering (VQA) pairs, and 3k dynamic object grounding samples.We comprehensively study general-purpose, spatial-aware, and region-level MLLMs to understand how they "think in dynamics" from both linguistic and visual perspectives. Our results reveal that existing models struggle to jointly excel in both spatio-temporal reasoning and dynamic object grounding, often producing inconsistent interpretations of motion and interaction. Conventional prompting strategies i.e., chain-of-thought or caption-based hints) provide only limited improvements.In contrast, structured integration approaches, including Mask-Guided Fusion and the Spatio-Temporal Textual Cognitive Map (ST-TCM), substantially enhance MLLMs' dynamic perception and spatio-temporal reasoning in an evolving 4D world. These findings underscore the importance of explicit spatio-temporal structural cues to bridge the gap between static perception and dynamic reasoning in MLLMs.