Recent advancements in unified generative models (UGMs) and world simulators have achieved unprecedented results in visual perception and synthesis. However, these models primarily rely on surface-level event alignment, leaving the capacity for high-level visual reasoning underexplored. True visual generative intelligence demands "Reasoning-to-Generation", an ability to infer latent rules from visual inputs and manifest solutions through precise, logically constrained visual outcomes. We introduce RIG-BENCH, a novel comprehensive benchmark that systematically evaluates Reasoning-driven Image Generation (RIG) across four cognitively demanding domains: Concept-based, Transformation-based, Pattern & Structure, and Scenario-based. Featuring 2000 curated samples, RIG-BENCH serves as a rigorous stress test for RIG. Our extensive evaluations of state-of-the-art UGMs and image/video generation models reveal a significant reasoning-generation gap, wherein models frequently produce locally plausible but globally illogical outputs. RIG-BENCH provides a vital diagnostic framework to guide the development of next-generation, logically grounded UGMs and world simulators.
External text can override conflicting image evidence in multimodal large language models, a failure we call multimodal contextual sycophancy. We introduce a 998-case diagnostic that independently varies visual evidence, commonsense priors, and external text, and probe when this failure arises by moving the information boundary around a context-blind visual witness. On abnormal images paired with Gemini-generated false text, GPT-5.1 scores 7.9% under joint conditioning, 49.7% when the context-blind witness report is scored directly, 63.7% under a matched two-call witness-arbiter pipeline that exposes the witness to the text, and 84.2% under System-2 Visual Arbitration (S2VA), which withholds the text from the witness. Across six models, S2VA improves over the direct witness report by 19.7 to 44.1 points, with all paired 95% confidence intervals excluding zero. The best information boundary is not uniform: textual context scaffolds some models, and a GPT-4o-regenerated subset changes the relative ordering of joint conditioning, Witness-Only, and S2VA. Contextual sycophancy is therefore sensitive to when text is introduced, as well as to the model and context source.
Tool-augmented multimodal reasoning integrates external tools (e.g., object detection, depth estimation) into multimodal large language models (MLLMs) to address perceptual bottlenecks in complex visual tasks. However, existing approaches rarely verify tool outputs, limiting their ability to detect and recover from tool failures. We propose ReVISE, a framework that equips MLLMs with verification and dynamic error recovery for tool-augmented reasoning. ReVISE introduces (1) a curated training dataset that supervises reflective behaviors, enabling models to validate tool-derived evidence, reformulate queries when visual mismatches arise, and fall back to intrinsic grounding when external tools are unreliable; and (2) a reinforcement learning based targeted rewards that encourage internal reflection and penalize spatial misalignment. Experiments on several benchmarks demonstrate consistent improvements over existing methods, highlighting the importance of error detection and correction in tool-augmented multimodal reasoning.
Chain-of-thought (CoT) reasoning has dramatically improved large language models (LLMs) by allowing them to decompose problems into intermediate steps. While CoT is widely effective for linguistic tasks, text-only CoT forces models to serialize visual problems into awkward prose. Although architectural solutions exist to process visual inputs, the community lacks a massive, multi-step, self-corrected dataset to teach models how to build and maintain internal visual workspaces when solving purely textual reasoning problems. To address this limitation, we introduce CoVA-SFT, a highly structured corpus of 51.9K samples containing over 222K multimodal reasoning steps across 5 distinct layout families and 17 complex tasks, and CoVA-Bench, a companion benchmark of 1,700 held-out test samples spanning the same tasks for reproducible evaluation. By providing explicit rationale formulations, agentic renderings, and verification loops, CoVA-SFT teaches multimodal language models to interleave text and visual abstractions. We validate the dataset by demonstrating that models fine-tuned on CoVA-SFT outperform all interleaved CoT baselines by more than 2x on average on CoVA-Bench, though they still fall short of strong text-only CoT baselines, highlighting open challenges for future work.
Native visual reasoning treats visual generation as the medium of reasoning itself: visual states (i.e. images and videos) are not merely inputs to be understood or outputs to be rendered, but first-class substrates for problem solving beyond language. Yet progress remains bottlenecked by the lack of scalable training tasks, reliable feedback, and controlled comparisons across generative substrates. In this work, we introduce VBVR-Pro, a closed-loop testbed that makes native visual reasoning through generation trainable, verifiable, optimizable, and experimentally controllable. 1) Task scaling. VBVR-Pro turns visual reasoning into a controlled task space of 300 procedurally generated tasks. Models trained on VBVR-Pro show strong transfer beyond the proposed suite across seven external visual reasoning benchmarks such as RISE-Video, MME-CoF-Pro, and BabyVision. 2) Verifiable rewards. VBVR-Pro provides verifiable reward scorers for task-grounded evaluation. Through a systematic study of leading MLLMs as judges, we identify recurring failure modes of the prevalent VLM-as-a-judge paradigm. In contrast, the proposed scorers are grounded in deterministic, task-specific rules, achieve fine-grained alignment with human judgments. Importantly, they serve as reliable reward signals for large-scale multi-task reinforcement learning and demonstrate stronger post-RL performance across visual reasoning tasks. 3) Mechanism study. VBVR-Pro enables controlled modality studies across more than 30 image, video, and interleaved generators. Our analysis shows that video generation remains strongest for tasks requiring persistent spatiotemporal state tracking, while interleaved generation provides a compute-efficient alternative. Critically, ablations and probing suggest the presence of vision-native trajectories that are crucial to visual reasoning. We release all data, models, scorers, and code.
Matthew Perlman, James Beetham, Niels Da Vitoria Lobo +2cs.CV cs.GR
Scalable Vector Graphics (SVGs) power much of the modern visual ecosystem, yet state-of-the-art generative models focus almost entirely on rasterized images. We explore whether inference-time methods can unlock SVG generation capabilities in off-the-shelf vision-language models (VLMs). We systematically evaluate a constrained iterative refinement harness that combines visual feedback, structured editing, and constrained decoding to characterize the capabilities and limitations of current VLMs for SVG generation. Across multiple VLMs and generation settings, we find that constrained decoding improves compilation success rates, while iterative refinement reveals a deficit in visual reasoning and self-correction. Our results highlight both the promise and current limitations of using inference-time methods to adapt general-purpose VLMs for SVG generation.
Vision-Language Models (VLMs) achieve strong performance in visual reasoning tasks, but it remains unclear whether they understand visual relations, or simply employ shortcuts such as language cues or priors. To investigate this, we use the Qwen3-VL-4B (Bai et al., 2025), a modern VLM, to decode how visual information is encoded across depths. For this, we propose a synthetic dataset of simple geometric shapes for controlled analysis, along with queries crafted to precisely test language cues. Furthermore, the dataset is modified to test causal reliance on visual evidence. Our results show that current VLMs combine genuine visual reasoning with shortcut strategies primarily rooted in language cues.
Two-stage neuro-symbolic architectures provide an elegant paradigm for visual problem solving by cleanly separating connectionist perception of predefined symbols from possibly later defined relational reasoning thereon. However, anchoring high-level predicates into visual frames typically necessitates annotations that are expensive to acquire. In this work, we introduce the Dynamic Orthogonal Concept Bottleneck (D-OCB), an object-centric slot- VAE framework designed to extract human-aligned symbolic predicates under extremely weak supervision. D-OCB eliminates the arduous manual tuning of loss-balancing coef- ficients by dynamically learning optimal hyperparameter allocations during training. To infuse prior knowledge on independence of concept categories, in addition to standard re- construction self-supervision we penalize correlation across concept subspaces. Crucially, to combat the instability of very low supervision regimes, D-OCB incorporates a dynamic di- mensionality allocation mechanism; this adaptive formulation allows well-represented con- cepts to yield latent dimensions to underperforming concepts that are lagging behind, effectively preventing representation collapse and significantly improving overall concept accuracy. Through an extensive empirical evaluation, we demonstrate that our framework achieves high concept alignment and downstream visual reasoning accuracy using minimal label budgets, matching or outperforming end-to-end paradigms.
Real-world situation appearances can deviate from their underlying physical states, challenging the reliability of multimodal large language models (MLLMs) in practical applications. In this paper, we term this phenomenon situational illusions and investigate: (1) how MLLMs perform under such illusions, and (2) how to mitigate the limitations. We first develop a comprehensive where-what-how taxonomy that characterizes where situational illusions occur, what targets they take, and how they arise. Building on this taxonomy, we introduce MSIBench, a benchmark designed to assess the discrimination, understanding, and reasoning capabilities of MLLMs under situational illusions. Evaluations of 27 model configurations reveal that current MLLMs are highly vulnerable to these illusions and exhibit 6 typical failure modes related to visual observation, grounding, and reasoning. To mitigate the limitations, we build on the core idea of systematically inspecting and reasoning over visual evidence for contextual understanding, developing prompting for closed-source models and supervised fine-tuning for open-source models, respectively. These two simple yet effective methods improve model performances by 20% at most, suggesting a practical path toward more reliable multimodal perception and reasoning in complex real-world environments.
Recent studies suggest that video generation models can exhibit certain forms of zero-shot visual reasoning through generated frames. Yet reliable evaluation remains challenging: benchmarks should adopt inputs aligned with the visual priors of current video models, require valid evolving processes rather than only plausible final states, and calibrate task difficulty to remain challenging yet partly feasible. To this end, we introduce VGI-bench, containing 27 tasks and 810 instances, organized by a two-level taxonomy of task domains and skill tags for fine-grained evaluation of visual reasoning capabilities of video generation models. Our evaluations show that current generative systems can solve a subset of visually grounded reasoning tasks, but remain far from reliable, with even the strongest model, Seedance~2.0, achieving only 51.0% under our evaluation criteria. Our analysis further explore the output failure modes, input condition sensitivity, performance transfer boundary from synthetic fine-tuning, and internal denoising perspective revealing limited self-correction, where later steps mainly refine early hypotheses rather than correct reasoning errors. We hope VGI-bench will help stimulate the development of next-generation video generation models. We will release our code and data.
Chart editing requires inferring and modifying visualization code from a reference chart image based on an editing instruction, challenging fine-grained visual reasoning, instruction following, and executable code synthesis capabilities of MLLMs. Large reasoning models (LRMs) with extended Chain-of-Thought (CoT) reasoning are suitable for tackling such complex multimodal tasks. However, our preliminary study reveals an ``inverted-U'' relationship between reasoning length and chart-editing performance: Excessive reasoning often leads to ``overthinking,'' where models drift toward hallucinated visual details or get stuck in redundant reasoning loops. To address the gap, we introduce REChart, a two-stage training framework that provides process-level supervision over intermediate reasoning steps, improving both editing fidelity and reasoning efficiency. First, we synthesize 200k high-quality reasoning trajectories for supervised fine-tuning from a large image-instruction-code pool, using a role-specialized agentic Reason-Score-Refine workflow that iteratively refine the chart code toward higher quality. Second, we optimize the model via reinforcement learning with two complementary rewards: a \emph{fidelity} reward evaluating code correctness, visual fidelity, and structural consistency, and an \emph{efficiency} reward that assigns each rollout a random thinking budget, truncates the reasoning process, and credits the final reasoning segment according to its contribution to the output. On the ChartEdit and ChartMIMIC benchmarks, our model achieves state-of-the-art chart-editing performance among open-source models of comparable scale, while mitigating overthinking and reducing average reasoning token usage by 79.0\% under a maximum thinking budget of 16,384 tokens compared with the base model.
A benchmark should deliver more than a scalar score: what makes an evaluation trustworthy is the reasoning that justifies the score. This is especially critical for world models, where judging a rollout requires understanding whether physics, causality, and world state evolve correctly. Humans spot such violations naturally, yet no existing benchmark automates this capability: metrics are computed brute-force, leaving no reasoning chain that can be examined or verified. We introduce HarnessEval-W, an agentified evaluation pipeline that brings the harness paradigm from the LLM ecosystem to world model benchmarking. Rather than applying a fixed rubric, HarnessEval-W interprets the context of each evaluation case, decomposes the evaluation question into measurable subproblems, and spawns specialized sub-agents, each equipped with tailored context and diagnostic tools to reason over its own subproblem. The parent agent then validates the gathered evidence and summarizes it into the final verdict. This hierarchical workflow turns every evaluation into a transparent evidence tree whose complete reasoning chain justifies the result. We apply HarnessEval-W to 18 representative world models over 330 evaluation cases. Its judgments closely align with human preferences while providing verifiable, fine-grained diagnoses of every generated rollout. We open-source the full pipeline as a live benchmark and invite the broad community to contribute to grow new skills and evaluation cases as world models evolve.
Chart-to-code generation requires a model to read the fine-grained visual details of a chart and write executable code that reproduces it. Existing chart-to-code methods either train visual and coding abilities separately, or fine-tune on chart-to-code data with the two abilities entangled. Neither strategy accounts for the distinct nature of the two abilities or the interference that arises when they are optimized together. We propose MoCA (Mixture of Cross-modal Arbitration), which separates the two abilities rather than blending them. MoCA is built on Cross-modal Arbitration Block (CAB), which maintains a visual branch and a code branch as two distinct pathways, and a lightweight arbiter that arbitrates their relative contributions at every layer and generated token. We train MoCA in two stages: a supervised warm-up on self-distilled reasoning trajectories that decomposes visual understanding into explicit steps, followed by reinforcement learning with rewards on both the reasoning process and the final code. Analysis shows that the arbiter learns structured rather than arbitrary allocations, with expert contributions varying systematically across tokens, layers, and instances. Across three benchmarks, MoCA delivers competitive performance against general-domain and chart-specialized models. Ablation results show that the gains cannot be attributed to a larger model size alone, but instead arise from the joint contributions of complementary visual and code branch initialization and input-conditioned arbitration through CAB.
Vision-language models are increasingly used for multimodal question answering, yet their ability to reconstruct latent spatial structure from a single image remains difficult to isolate. Broad benchmarks often combine perception, optical character recognition, domain knowledge, linguistic priors, and reasoning in the same evaluation. We introduce StateSight, a procedurally generated benchmark for cube-net opposite-face reasoning, occluded cube-tower counting, and 4-neighbor connected-component counting. Each task family contains 300 single-image prompts with deterministic oracle labels and exact-match scoring. OpenAI GPT-5.5, using the API model identifier gpt-5.5, achieved 59.3%, 33.3%, and 28.3% accuracy across the three tasks, while Claude Sonnet 5 achieved 53.3%, 18.7%, and 7.3%. All final direct runs had zero format errors. A 30-participant human baseline on 60 items exceeded both models on every task, with mean accuracies of 80.8%, 68.8%, and 64.3%. Visible-derivation analysis identified recurring errors in image-state reconstruction and reasoning procedure. We also introduce StateSight-Steps, a companion dataset of 900 interleaved image-text examples and 3,600 deterministic intermediate visual states. The results show that format-valid responses can mask failures to recover the spatial structure required for verifiable visual inference.
Multimodal large language models (MLLMs) exhibit substantial performance degradation in non-English visual reasoning, despite the strong multilingual competence of their text-only backbones. While mechanistic evidence from text-only models suggests that non-English inputs are routed through an English-centric latent space, the multimodal implications of this phenomenon remain unexplored. Through rigorous mechanistic analysis, we identify the \textbf{Ghost Anchor} phenomenon: a temporal modality asynchrony where linguistic translation to the English semantic manifold completes in early layers, while visual semanticization remains immature. Consequently, visual signals are physically present yet functionally invisible during the early alignment window. To rectify this, we propose \textbf{ANCHOR}, a training framework employing Proactive Visual Anchoring (PVA) to accelerate early visual semantic emergence, ensuring visual representations proactively guide linguistic translation. Mechanistic interventions confirm that ANCHOR successfully restores the causal influence of visual signals during early translation. Furthermore, extensive experiments on XMMMU, MaXM, and CVQA demonstrate that ANCHOR consistently outperforms standard baselines, achieving robust visual reasoning across both fine-tuned and zero-shot languages.
Although visual reasoning is crucial for solving complex geometry tasks, existing vision-language models rely heavily on text-only reasoning. Some recent methods introduce intermediate visual states to facilitate reasoning, but they are often hindered by inaccurate geometric representations and low rendering fidelity, ultimately leading to unreliable outputs. To address these limitations, we propose MetaReason, a framework for multimodal reasoning in plane geometry that leverages structured meta-information to enable accurate auxiliary-line construction. The framework first parses geometric images into meta-information, performs controllable edits with predefined tools to synthesize high-fidelity visual states, and then conducts reasoning based on these augmented views. To support this framework, we construct TutorGeo, a comprehensive dataset containing 17k image-to-meta conversion samples, 60k text-only reasoning traces, and 60k interleaved multimodal reasoning traces. Using this dataset, we combine supervised fine-tuning and reinforcement learning to develop robust multimodal reasoning capabilities. We also introduce ExamGeo, a benchmark derived from real-world examination problems that enables systematic evaluation across varying difficulty levels. Experimental results demonstrate that MetaReason significantly outperforms existing open-source models and achieves competitive performance against proprietary models.
Cloud-hosted vision-language models (VLMs) offer greater contextual reasoning capabilities than smaller onboard models, but frequent visual uploads increase communication overhead and add network and inference latency to tactical decisions. We present a risk-adaptive edge-cloud architecture in which onboard traffic assessment determines when cloud reasoning is requested. An onboard VLM and a lightweight detector capture temporal traffic conditions and path-relative hazards for conservative local response and selective cloud access. The cloud model provides tactical advice, while validation, vehicle control, and automatic emergency braking remain local. In CARLA experiments, our method matched the task success rate of periodic cloud access while reducing cloud requests by 54.1% and recording fewer automatic emergency braking (AEB) activations. In a delayed-roadwork ablation, semantic events triggered requests before the next scheduled audit. Across three emulated network profiles, the method continued to reduce cloud traffic, although lane changes took longer than with periodic access. Onboard traffic assessment therefore served as a practical trigger for selective VLM inference in these experiments.
Vision-language models (VLMs) remain unreliable on chart questions that require reasoning over visual quantities, and this weakness is usually attributed to a reasoning deficit and addressed with more reasoning supervision. We ask whether the difficulty lies in reasoning itself, or in the simpler skills that reasoning operates on: reading the plotted elements (\emph{perception}), locating them and binding them to their labels (\emph{grounding}), and performing single-step computations such as ranking, totals, and differences (\emph{simple reasoning}). We introduce \textbf{ChartProbe}, a diagnostic framework whose probes are generated directly from the code that renders each chart, so every gold answer is exact by construction, needs no human annotation, and attributes each failure to a single skill. ChartProbe enables an intervention prior work does not attempt: instead of synthesizing complex-reasoning data, we withhold complex questions and reasoning traces entirely, fine-tune on one simple skill at a time, and measure transfer to held-out complex-reasoning questions. Across three open-weight VLMs, supervising the simpler skills alone produces large gains on complex-reasoning questions the model never trained on: where these skills are weak and the model can be taught to read the image, training them recovers much of complex reasoning at no reasoning-data cost. The gains hold across three out-of-distribution settings: an unseen chart type (pie charts), a human-written benchmark disjoint from our images and templates (ChartQA), and a non-chart visual domain (CLEVR). Complex visual reasoning can therefore improve without complex-reasoning supervision.
Vision-language models (VLMs) have been widely employed in understanding and reasoning tasks for unmanned aerial vehicles (UAVs). Existing UAV benchmarks primarily focus on aerial-view scenarios. However, whether current VLMs can perform well on understanding and reasoning tasks in aerial-ground collaborative scenarios which are practical in real-world applications like rescue and infrastructure inspection remains underexplored. To address this gap, we introduce AeroGround, a comprehensive benchmark for evaluating VLMs in aerial-ground collaborative reasoning. AeroGround is built upon a simulated aerial-ground dataset containing approximately 29,000 multimodal observation groups from diverse open environments, and provides 2,250 high-quality question-answering instances covering cross-view correspondence, spatial understanding, and reasoning. Experiments on 16 pretrained VLMs, together with two domain-adapted variants, reveal a substantial gap between current models and human performance: the best model achieves an average accuracy of 54.4%, whereas humans reach 93.3%. By systematically revealing the strengths and limitations of existing models in aerial-ground collaborative reasoning, AeroGround provides a foundation for developing more capable aerial-ground collaborative embodied intelligence systems.
Multimodal models increasingly reach for tools when solving visual tasks (crop, zoom, rotate, brighten), a paradigm known as thinking-with-images. The central challenge is one of perception: tools mostly serve to expose visual evidence, reasoning over that evidence stays in language, and most targets are ones a human could in principle determine by inspection. Some visual questions, however, are not bottlenecked by perception: recovering their answers requires executing a multi-step visual algorithm over the pixels. On such questions a model often names the correct algorithm at once yet still answers wrong, because language can describe an algorithm without being able to run one. Code-with-Image crosses that line: given nothing but a Python interpreter, the model must implement a genuine visual algorithm in code to solve the task; the program itself becomes the reasoning. The bottleneck then shifts from executing code to deciding which algorithm to implement. So we let the model teach itself: a training-free reflection loop studies its own failed programs, tests repairs against constructive ground truth, and keeps what survives as portable skills. On our Code-with-Image Bench (CwI-Bench), thirty task families induced by hidden visual computations with disjoint learning and evaluation splits, even GPT-5.6-luna stays below 30% with tool-free chain of thought; given a bare interpreter it reaches 43%, and with skills evolved through its own executable reflection, 67%. The open 27B model climbs the same ladder (9% $\rightarrow$ 33% $\rightarrow$ 56%), and the skills are plain text, transferable across scales and families. When code carries the reasoning, debugging code becomes debugging reasoning.
Medical visual agents can use tools to inspect images and retrieve external knowledge, but indiscriminate tool use may introduce noisy or misleading evidence. Reliable diagnosis therefore requires not only acquiring additional observations, but also verifying whether tool actions are necessary and whether the resulting evidence supports the current hypothesis. We introduce MIRA (Medical Image Reflection for Agentic Diagnosis), a medical visual diagnostic framework for autonomous evidence search and reflective verification. MIRA dynamically invokes image-processing operations, including zooming, grounding, pointing, rotation, and measurement, as well as web search, while evaluating the relevance and consistency of the acquired evidence. We develop MIRA through a two-stage training strategy. First, a tool-augmented Monte Carlo Tree Search data engine explores diverse diagnostic hypotheses and jointly verifies visual grounding accuracy and semantic consistency to construct supervised fine-tuning trajectories. Second, reinforcement learning further improves decision-making through online reflective principle evolution: failure cases are distilled into candidate principles, and only principles that improve held-out rollout rewards are retained. Across nine medical visual reasoning benchmarks, MIRA achieves an average score of 64.73, improving its Qwen3-VL-8B backbone by 7.44 points. It also increases useful tool-use judgments from 56.2% to 73.8% and reduces harmful judgments from 8.9% to 1.6%. Qualitative analyses show that MIRA can re-examine evidence, correct premature conclusions, and adapt its tool-use strategy. Project page: https://MIRA-VL.github.io/
Tool-augmented vision-language models increasingly "think with images": they call crop, zoom, or code tools and reason over the returned pixels. However, recent work using blind tests, gain decompositions, and attention analyses has shown that returned images contribute little, raising the question: if pixels do not carry the gain, what does? We hypothesize that the load-bearing signal is the structured text emitted before any returned pixel arrives: tool name, coordinates, target description, and intent. This textual scaffold encodes where to look and what to find. We introduce TextCall (call-but-no-return) to test this: it keeps the scaffold but replaces returned images with the text placeholder [Image output skipped]. Three studies support the hypothesis. (i) Non-necessity of returned pixels: across LoRA, full fine-tuning, and RL, TextCall matches or exceeds full thinking-with-images; under RL it preserves tool use at the reported checkpoint, avoiding the failure mode where, under matched settings, seeing the returned image causes the model to stop calling tools and answer directly. (ii) Sufficiency of the scaffold: on matched training queries, scaffold-only input yields equivalent accuracy to returned-image input. (iii) Component specificity: decomposing the scaffold into reasoning text and spatial code shows both components contribute, with the dominant one varying by task. Together these results support the Tool-Call Scaffold Hypothesis: in current thinking-with-images distributions, the active signal is the structured text emitted at tool-call time; the returned image is a redundant carrier. TextCall preserves accuracy while reducing latency by 29-46% and eliminating tool-execution API calls. Our claims hold for current thinking-with-images benchmarks; constructing tasks where pixels are genuinely load-bearing remains an open direction.
Electrical circuit analysis requires more than recognizing components in an image. A solver must ground symbols and labels, recover latent topology, select a physical model, formulate coupled equations, propagate intermediate quantities, and preserve units, signs, directions, and phase conventions. We introduce \benchmark, a benchmark of 1,000 authentic textbook problems for evaluating this complete long-horizon visual-to-symbolic reasoning process. Each problem pairs one or more circuit diagrams with a self-contained question, a typed or semantically specified answer, and a reference worked solution. An evidence-first construction pipeline aligns questions, figures, and solutions, while a reasoning-oriented taxonomy organizes problems by circuit type and dependency depth. Evaluation combines conservative typed scoring with identity-blinded multi-model semantic consensus, retaining every problem in the denominator. Across three commercial chatbot systems and six open-source multimodal large language models, the highest-scoring system reaches 84.8\% accuracy. However, performance consistently deteriorates on long-horizon problems, and qualitative analysis exposes persistent failures in topology-to-target binding, physical conventions, and late-stage output propagation. \benchmark{} provides a focused testbed for measuring whether multimodal models can transform technical visual evidence into sustained, physically valid symbolic reasoning. Code are available at GitHub - CircuitReason/CircuitReason1K.
Industrial-safety understanding requires more than detecting workers, equipment, and personal protective equipment. Models must also assess compliance, identify hazardous interactions, explain potential accident mechanisms, and recommend preventive actions. Existing safety datasets primarily focus on visual perception or isolated violation recognition and provide limited supervision for evidence-grounded reasoning. We introduce SafeSceneReason, a multimodal industrial-safety reasoning benchmark and companion training corpus that connects workplace scenes with knowledge from occupational accident investigations. SafeSceneReason combines two complementary data-construction pipelines. The scene-centric pipeline converts annotated workplace images into executable safety scene graphs and generates deterministic answers through program execution over objects, relations, and safety rules. The report-centric pipeline extracts figures and contextual evidence from accident reports and constructs multimodal questions using evidence graphs, explicit information boundaries, multi-step reasoning paths, and iterative verification. The resulting resource contains 110,581 verified scene-centric question--answer pairs and 13,114 refined report-centric question--answer pairs, covering perception, spatial and quantitative reasoning, compliance assessment, evidence synthesis, causal analysis, and mitigation-oriented decision making. Evaluation of representative proprietary and open-source vision--language models reveals substantial performance differences and persistent weaknesses in comparative, technical, and multi-evidence reasoning, demonstrating that strong general visual understanding does not yet guarantee reliable industrial-safety reasoning.
Vision-Language Models (VLMs) should answer from concrete image evidence rather than language priors, dataset shortcuts, or irrelevant visual context. Existing perception-aware post-training methods encourage image use through global perturbations or attention proxies, but they do not test whether a sampled answer causally depends on the local evidence that supports it. We propose Counterfactual Evidence Disentanglement (CED), a training-time evidence audit for VLM grounding. For each response, CED neutralizes an object-centric Evidence Region and compares the resulting support drop against matched non-evidence Regions. We combine this signal with answer correctness inside GRPO, rewarding correct answers that rely on the evidence path rather than shortcut or nuisance paths. CED uses weak object-level proposals, requires no question-specific evidence annotations, and adds no inference-time overhead. Across nine public benchmarks and four backbones, CED outperforms prior RL-based post-training methods, with targeted analyses verifying its object-centric signal.
The visual reasoning ability of multimodal large language models (MLLMs) is crucial for downstream applications, particularly counter-commonsense reasoning, which requires models to reason beyond common assumptions. Recent studies mainly improve visual counter-commonsense reasoning by enhancing visual inputs, following the assumption that failures originate from insufficient visual grounding. However, our empirical analysis reveals that the bottleneck is not visual perception. MLLMs already capture the relevant visual evidence, and the correct answer exists in their decoding space. Instead, the shared language decoder resolves prior--evidence conflicts by favoring dominant language priors, especially for low-frequency factual scenarios. Motivated by this, we first propose a text-anchored data construction pipeline, whose core component, Fact-Frequency Distillation (FFD), estimates the prior strength of commonsense facts and distills verified counter-commonsense scenarios into a high-quality text corpus. Building upon this corpus, we introduce TACT, a text-anchored post-training framework that debiases the shared language decoder without requiring any visual training data. TACT routes evidence-following and prior-driven reasoning trajectories into different optimization stages, enabling the decoder to resolve prior--evidence conflicts. Across counter-commonsense visual benchmarks, TACT substantially improves visual reasoning while preserving general capabilities, demonstrating effective text-to-vision cross-modal transfer.
On-Policy Self-Distillation (OPSD) has become a standard post-training approach for improving visual reasoning in multimodal large language models (MLLMs). Existing methods draw privileged information from diverse input sources to guide self-distillation. Yet these designs overlook Modality Imbalance, a challenge inherent to MLLM reasoning. When textual information dominates generation, the model cannot fully integrate its multimodal input. Consequently, carefully designed privileged information remains underused, limiting the effectiveness of OPSD. To examine this limitation, we construct a Positive Teacher with the Zoom-In Image and a Negative Teacher with the Mask Image, which exhibit different degrees of Modality Imbalance. Changes in their reasoning correctness and token logits reveal that Modality Balance can itself serve as privileged information. Motivated by this finding, we introduce OPD-V, a visual OPSD paradigm that instantiates such information through the Positive Teacher and Negative Teacher. Positive Modality-Balance Logits Margins define a Modality-Balance Trust Region that selects the on-policy tokens used for self-distillation. Experiments across 6 benchmarks, 4 MLLM backbones, and 5 post-training methods show that OPD-V consistently improves reasoning performance while reducing training cost.
Tianbao Jiang, Weicong Ni, Gerard de Melo +1cs.CL cs.AI
Post-training reinforcement learning (RL) algorithms are commonly used to align large vision-language models (LVLMs) with human intent and the requirements of visual reasoning tasks. However, existing RL-based alignment methods are often resource-intensive and encounter mismatches between training objectives and inference-time distributions. To bridge this gap, we propose a novel test-time alignment approach that leverages trajectory-guided structured sampling for dynamic inference-time refinement, achieving better alignment with visual grounding and ensuring logical consistency. Our approach begins with curating a reasoning memory bank via a trajectory learning algorithm, which decomposes complex question solving into ordered sequences of predefined reasoning patterns. It subsequently accomplishes inference-time alignment by first collecting trajectories from reasoning memory bank to establish a global structural reasoning prior, and then using an iterative Markov Chain Monte Carlo (MCMC) algorithm for localized multi-objective refinement of the reasoning trace. Experiments across multiple multimodal reasoning datasets demonstrate that our approach significantly improves accuracy without incurring prohibitive inference overhead. These results establish trajectory-guided test-time sampling as a scalable and effective alternative to traditional post-training alignment, particularly for complex visual reasoning tasks.
Vision-language models (VLMs) are expected to revise their reasoning when visual evidence changes. Failures to do so are often attributed to insufficient visual attention or contextual inertia, leaving unclear what models reuse instead of recomputing from the current image. We show that evidence-bearing reasoning in a prior chain of thought (CoT) can form a textual shortcut that competes behaviorally with visual recomputation. Across 16 VLMs, a matched counterfactual analysis identifies evidence-bearing content as the most robust carrier of prior-CoT influence. Removing this evidence-bearing content shifts answer preference more than removing length-matched non-evidence context or the final-answer span, with prior control weakening progressively as more stale evidence is removed. Reordering this evidence also weakens prior control, showing that its organization modulates shortcut strength. Beyond the immediate answer, the shortcut can retain residual influence after answer correction: weakening current-image support shifts preference back toward the prior answer, while repeated prior answers and reused premises arise mainly when the shortcut remains active. To limit this influence, we introduce Fresh-State Attention Firewall (FSAF), a training-free intervention that isolates fresh computation from the prior CoT. Across five VLMs, FSAF raises visual update rate from 35.28% to 53.61% and reduces prior-answer rate from 39.22% to 3.67%. Reliable VLM self-reflection therefore requires more than looking again: fresh visual recomputation must be protected from stale textual reuse.
Multimodal Large Reasoning Models (MLRMs) have achieved strong performance on tasks requiring visual understanding and multi-step inference. However, as reasoning trajectories grow, models may become less effective at using information established earlier in the context, increasing the risk of reasoning errors. Existing approaches primarily address this problem by sustaining visual grounding throughout reasoning. However, reasoning also transforms visual observations into task-specific relations, constraints, and intermediate conclusions whose influence may weaken over long trajectories. Our attribution analysis suggests that correctness is not consistently separated by image attribution alone, but is more closely associated with whether trajectories retain and integrate such reasoning-derived information across stages. Motivated by this, we introduce TRAM (TRajectory-derived Auxiliary Memory), a training-free method that augments standard decoding with an auxiliary memory pathway derived from the model's own reasoning trajectory. TRAM consolidates completed reasoning into a compact latent memory, updates it online through fast and slow recurrent streams, and feeds it back into selected decoder layers through a lightweight residual pathway. Experiments across four MLRM variants on eight benchmarks show that TRAM improves performance over vanilla decoding on mathematical, scientific, and general visual reasoning tasks without additional training.