VLM Reasoning

Multimodal large language models (MLLMs) struggle with spatial, embodied, physical, and temporal reasoning. We hypothesize that these failures reflect a shared deficit in visual transition reasoning, and test whether this capability can serve as a shared training primitive, one that different models can learn from different supervision sources and reuse across different tasks, with a systematic training recipe. Existing benchmarks document these deficits separately but do not support controlled comparisons across scenes, actions, and reasoning operations. We therefore introduce WOVEN, a training source and benchmark for visual transition reasoning that organizes transition supervision by scene, action, and reasoning type, using diverse, realistic rollouts from video-pretrained generative models: 36,076 examples across 20 scene types, 5 action types, and 8 reasoning types. We first evaluate 38 frontier MLLMs (e.g., GPT-5.4 and Qwen3-VL-235B-A22B) and find a substantial and systematic deficit: even the strongest models fall far below humans, and the failures recur across model families and persist with scale. We then train MLLMs at multiple scales on WOVEN and find that they learn a shared capability that transfers broadly: training subsets of only about 2,000 items each collectively improve 22 of 26 external benchmarks by up to 27.3 percentage points, and WOVEN data can replace 30-50% of a task's own training data with comparable accuracy. Controlled comparisons further yield a training recipe for visual world modeling, validated prospectively on held-out benchmarks: select supervision by the reasoning operation it teaches rather than by the actions, scenes, or domains it shows, and prefer larger changes to the visual state for robustness. Our work establishes visual transition reasoning as a reusable foundation for systematic visual world-model training in MLLMs.

Existing spatial reasoning benchmarks mainly test spatial perception: reading off relations already visible in the input. Yet real-world spatial intelligence demands predictive spatial reasoning: constructing a scene from observations, anticipating how an intervention changes it, and reasoning about the unseen outcome. We introduce SpaceCast-Bench, the first benchmark to directly and diagnostically evaluate this capability. Built around an observe-transform-infer framework, its 3,862 questions from 182 real-world scenes span 16 task types at three levels: static perception, local prediction, and global prediction, progressively requiring scene understanding, spatial state updating, and relational inference over unobserved outcomes. Evaluating 21 models exposes a stark gap: the strongest model reaches only 58.0% against 87.2% human performance, while spatially specialized models remain near random chance. Controlled analyses further reveal that bridge views are critical for integrating distributed observations, and that explicit 3D evidence benefits models more reliably than generated outcome images or videos. Fine-tuning on our programmatically generated data lifts Qwen3-VL-4B from 34.0% to 65.7% with macro-average gains across six out-of-domain benchmarks.

Spatial reasoning remains a persistent weakness of vision-language models (VLMs), because RGB inputs do not directly provide geometric evidence. Existing remedies either inject 3D into the model at inference, paying architecture and latency costs, or train with outcome rewards that supervise only the final answer. Spatial errors originate in perception: a misjudged depth or direction can be corrected only by the scene's true geometry, which the 3D-scanned sources of spatial training corpora already provide. We propose GPD (Geometry-Privileged Distillation), which makes geometric evidence the privilege in on-policy self-distillation (OPSD). For each question, depth, semantic, and bird's-eye-view (BEV) cues are rendered as compact text and routed to the teacher alongside the reference answer; a privileged KL, applied only to incorrect trajectories, augments GRPO, and the deployed model remains RGB-only. On the 4B backbone, GPD achieves 57.1 on VSI-Bench and 37.6 average across MindCube, SPARBench, MMSI-Bench, and ViewSpatial, outperforming both GRPO and answer-privileged OPSD across spatial reasoning benchmarks. Ablations confirm the complementarity of 3D and answer privilege, the advantage of question-conditioned routing over full-context injection, and the benefit of restricting distillation to incorrect trajectories.

Processing long visual token sequences from high-resolution images makes multi-step reasoning computationally expensive for multimodal Large Language Models (MLLMs). Existing one-shot pruning and aggregation methods compress visual tokens into a fixed context before decoding. However, visual evidence needs can shift as reasoning unfolds, making it difficult for a fixed compressed context to retain all the details needed across stages. To address this challenge, we propose ViMoD, a lightweight framework that maintains a compact visual context while preserving access to original fine-grained evidence as reasoning needs evolve. Deformable Aggregation of Region-wise Tokens (DART) learns content-adaptive groups and aggregation capacities, constructing compact Coarse representations linked to recoverable original Fine tokens. Temporal Routing for Adaptive Contextual Evidence (TRACE) integrates decoding history to anticipate upcoming evidence needs and select, retain, or replace active Fine-token groups. Selected Fine tokens augment the persistent Coarse context in the frozen backbone, enabling stage-specific evidence access without continuously attending to all visual tokens. On Qwen3-VL-4B, ViMoD outperforms all evaluated baselines on all eight reasoning benchmarks at a 20% target visual token budget, improving the mean normalized score by 39.0% over the strongest evaluated one-shot baseline. These gains are achieved with only 0.0546% additional trainable parameters relative to the frozen backbone.

Existing approaches to multi-view spatial reasoning operate largely on sparse input views. Vision-language models (VLMs) are thus restricted to understand a scene and infer spatial relations within these fixed views, leading to fragile cross-view alignment and geometry-to-language bottleneck. To address these issues, we formulate a novel Seek-and-View reasoning approach to find implicit cross-view spatial evidence by locating a question-relevant view to support the spatial reasoning. To realize this approach, we propose Vantage, a training-free model-agnostic reasoning framework that pairs a VLM with a 3D foundation model: a viewpoint-grounded reasoning stage for question analysis and view planning, followed by a geometry-grounded evidence augmentation stage to effectively synthesize and incorporate visual evidence into the final reasoning. Comprehensive experiments on six VLMs demonstrate consistent improvements on five benchmarks without fine-tuning. Overall, by revealing spatial evidence through view-grounded reasoning, Vantage can largely reduce reliance on language-based cross-view alignment and improve multi-view spatial understanding. Our code is available at https://github.com/q1xiangchen/Vantage.

Spatial coding agents significantly improve spatial reasoning in Multimodal Large Language Models (MLLMs) by using external tools to generate verified execution traces. However, this paradigm inherently suffers from prohibitive inference-time overhead and external dependencies. In this paper, we explore whether an MLLM can internalize this agentic capability to operate entirely tool-free. We begin with a simple observation: prompting an MLLM with summarized execution traces of a spatial coding agent naturally unlocks the model's internal spatial Chain-of-Thought (CoT). Motivated by this, we introduce SpatialOPSD, an on-policy self-distillation framework that internalizes spatial reasoning into a standalone MLLM by formulating verified agent traces as privileged information. To mitigate privileged-information leakage during distillation, we introduce Repetition-Aware Distillation, which combines repetition masking with unlikelihood regularization. Experiments across multiple benchmarks demonstrate that self-distilling SpatialOPSD achieves higher average accuracy than SFT and GRPO on both spatial and OOD datasets, exhibiting superior performance and generalization.

Video-language models commonly assume that more temporal observations lead to more reliable reasoning. We question this assumption and argue that the key challenge is not merely processing more video frames efficiently, but learning to reason reliably under limited temporal evidence. We propose SAVER, a dense-to-sparse post-training framework that uses dense video views as training-time references for sparse-frame inference. During reinforcement post-training, paired dense and sparse views are optimized with grounding rewards and a reliability-gated reference reward, encouraging sparse view predictions to preserve task-relevant temporal evidence. Notably, SAVER is trained only on 1,250 randomly sampled temporal grounding examples, without using any video question answering annotations. Across three temporal grounding benchmarks and six video question-answering benchmarks, SAVER consistently improves performance across frame budgets. In particular, SAVER can match or surpass dense-frame Qwen3.5 baselines while using substantially fewer frames. These results show that temporal grounding can serve as an effective evidence-localization proxy for learning sparse video reasoning that transfers to broader video understanding tasks.

Vision-Language (VL) reasoning requires a model to both extract relevant and accurate information from an image (visual reasoning, VR), and to infer the answer from it (language reasoning, LR). Reinforcement learning with verifiable rewards typically trains both through a single chain-of-thought with a final-answer reward. This gives every CoT token the same sequence-level advantage, failing to distinguish capability specific errors. We propose SPLIT-RL, a staged post-training approach that trains VR and LR in disjoint phases. Because a group's rollouts differ along one capability at a time, the group-relative advantage isolates it, and each phase is optimized using phase-specific reward. We further introduce Claim-Level Advantage (CLA-GRPO), which decomposes VR-phase rollouts into atomic visual claims and provides a fine-grained advantage at claim level based on visual-type group formation. Although trained in two phases, trained policy is evaluated like GRPO model, with a single CoT call at inference time. Under this protocol, SPLIT-RL improves average accuracy over GRPO by 1.4-6.1 points across Qwen3-VL models from 2B to 30B-A3B and InternVL3.5-8B. Evaluating each capability using an oracle based diagnostic shows that answer-only GRPO leaves perception unchanged, whereas SPLIT-RL improves both VR and LR.

Reinforcement learning with verifiable rewards (RLVR) has become a standard recipe for post-training vision-language models (VLMs), but it typically assumes a static training environment. As the actor improves, fixed tasks drift out of its learning frontier: many become trivial, others remain unsolvable; and the learning signal collapses. We argue that VLM post-training should evolve the visual environment alongside the actor, not just the actor itself. We propose VICO, a co-evolutionary framework in which an actor and an Environment-as-Rewriter (EnvRewriter) are trained jointly: the EnvRewriter edits verifiable image-side structures, such as scene graphs, chart tables, or protected region masks, and re-renders them to produce label-valid training samples whose difficulty is calibrated to the actor's current ability through a pass-rate-based reward. This loop continuously realigns task difficulty with actor capability without any additional human annotation. Across nine multimodal benchmarks spanning mathematical reasoning and visually grounded understanding, VICO-8B improves over its base model by up to +5.0% on out-of-domain tasks, surpasses the strongest self-evolution and text-editing co-evolution baselines by +4.3% and +8.4% respectively, and stays comparable to chart-specialized RLVR methods using 16-160 times fewer labeled samples. By shifting from human-labeled supervision to image-editing co-evolution, VICO offers a scalable path beyond static-corpus RLVR for visual reasoning.

Dual-process agents pair a fast policy with a slow deliberative model. In real-time settings the slow model usually runs continuously; in turn-based agents and robot planners it is invoked on events such as uncertainty or a detected failure. We study a fast learned actor that takes every decision and hands control to a reasoning vision-language model only when a gate opens, while the game keeps running. We use closed-loop Doom and the new open "System One" typed-decision models, served through a common llama.cpp interface. On 900 held-out questions, (i) zero-shot decision models from 0.15B to 9B parameters choose to collect items 1.6-1.8 times more often than chance among their errors, in any option order, although the order changes some models' accuracy; (ii) accuracy, calibration and sensitivity (how well confidence separates right from wrong answers) are distinct: models of similar accuracy differ widely in AUROC, and the confidence of the most sensitive one tracks which kinds of situation it fails, not which answers are wrong; (iii) offline, deferring the least confident 30% of decisions to a reasoning model gains over random deferral in proportion to the actor's AUROC (rank correlation 0.87); with actor and rate chosen on held-out games the gain is +0.13 [0.08, 0.18] with doomLaya's option order and +0.08 [0.02, 0.14] with shuffled options, and reasoning carries about half of it; (iv) in closed loop (33 games, three seeds) no variant reaches the exit. Committing to plans, the reasoner's or a fixed explore rule's, opens more doors and makes an actor that stands still play; with the rule the agent dies more often. Told that some doors need keys, the reasoner takes ordinary doors for locked ones, which the state cannot tell apart; without that knowledge it goes back to collecting. We release code, prompts, data and logs.

Pre-trained vision encoders contain layer-wise visual representations that differ in spatial granularity, semantic abstraction, and sensitivity to local details. However, most Multimodal Large Language Models (MLLMs) rely on only the final or penultimate vision encoder representations or fixed aggregation rules, making visual abstraction largely query-agnostic and limiting access to fine-grained cues such as small objects, spatial details, text, and subtle visual attributes. In this work, we propose Mixture of Layers (MoL), an instruction-conditioned layer routing approach at the visual patch level that dynamically aggregates query-relevant latent representations from intermediate vision encoder layers. Given a text query, MoL predicts routing probabilities over vision encoder layers and performs a top-k sparse aggregation over selected hidden states at either the image level, patch level, or through a hybrid routing mechanism. In doing so, MoL enables query-adaptive access to layer-specific visual features for fine-grained visual reasoning. Our experiments across 7 fine-grained visual reasoning tasks demonstrate substantial performance improvements, especially across fine-grained visual grounding and understanding tasks such as +18.9% improvement on V* in overall accuracy, +4.5% on HRBench4K, and +16.3% on CharXiv compared to the baseline MLLMs, without resorting to multi-resolution inputs, simple interleaving of multiple vision encoders, or increasing the number of patch tokens. We study vision encoders' receptive field scales across different layers and their sampling behaviors to provide an in-depth analysis of why layer-wise sampling is helpful, demonstrating that conditional visual representations are a key step towards better visual perception and reasoning in MLLMs. Our project page is available at https://wjdghks950.github.io/mol.github.io/.

Embodied agents must reason about 3D space while the video is still arriving, answering questions as soon as they have observed enough of the scene. VLMs that incorporate 3D geometric priors achieve strong spatial reasoning, but they operate offline, i.e., the full video must be available before they produce an answer. Streaming VLMs process frames causally and decide for themselves when to respond, yet they lack explicit 3D representations. We present SpaTime, a streaming VLM that fuses causal geometry tokens into the language model at every frame, using only the frames observed so far. To supervise when the model answers, we propose a response-time loss that maps per-frame response probabilities to a differentiable expected response time and penalizes the distance from the ground-truth frame. For evaluation, we construct StreamVSTI-Bench and StreamVSI-Bench, streaming adaptations of VSTI-Bench and VSI-Bench. On StreamVSTI-Bench, SpaTime reaches 49.2% overall accuracy and reduces the mean response-time error by 66% relative to the strongest streaming baseline.

Model merging provides a training-free way to transfer reasoning capabilities from language models to vision-language models (VLMs), but endpoint-based transfer can conflate pre-existing model differences with changes acquired during reasoning post-training. We instead formulate capability transfer around the training-stage update, isolating the parameter changes induced by reinforcement learning (RL). Yet transferring this update in full remains suboptimal: we find that its components differ substantially in cross-model transferability, with dominant directions transferring more effectively than the complete update. Based on this finding, we introduce Selective-RL, which isolates the RL-stage update, retains its dominant matrix-wise directions with magnitude preservation, and transfers them to the language modules of a VLM. Across three model families and five visual-reasoning benchmarks, Selective-RL improves full-update interpolation in 12 of 15 comparisons, including an 8.55 percentage-point MathVision gain on the Qwen recipient. Matched controls show that update magnitude or arbitrary low rank alone does not reproduce these gains. These results highlight a distinction between what is acquired during post-training and what remains transferable across models, providing a training-stage perspective on cross-model capability transfer. Code is available at https://anonymous.4open.science/r/selective-rl.

Assistance in collaborative manipulation is often initiated by user instructions, making high-level reasoning request-driven. In fluent human teamwork, however, partners often infer the next helpful step from the observed outcome of an action rather than waiting for instructions. Motivated by this, we investigate an event-driven formulation of proactive assistance, where human--object interaction outcomes initiate assistive reasoning without user-provided task specifications at inference time. To this end, we propose an event-driven framework that monitors workspace state changes with an event monitor and, upon event completion, extracts stabilized pre/post snapshots that characterize the resulting state transition. A frozen pretrained Vision-Language Model (VLM) then uses its semantic priors to infer the task context, decide whether assistance is appropriate, and, when needed, generate a sequence of assistive actions from the observed transition. To make outputs executable and verifiable, we restrict actions to a set of action primitives and reference objects via integer IDs.We evaluate the same framework across three distinct real world tabletop collaboration tasks without task-specific training or fine-tuning. The event-driven framework achieves performance comparable to variants given user instructions.

Thinking-enabled vision-language models (VLMs) report ever-higher accuracy on spatial benchmarks, yet final-answer scores cannot reveal whether a correct prediction reflects faithful spatial reasoning or a linguistic shortcut. We introduce SpatialChain, a dataset of 28,350 training and 899 test examples pairing spatially-oriented GQA questions with scene-graph-grounded reasoning chains, retained only when the generated answer matches the symbolic ground truth, and a two-axis evaluation combining objective chain-overlap metrics with a scene-graph-aware LLM judge that scores faithfulness and completeness independently of the final answer. Applied to nine thinking-enabled VLMs, the protocol surfaces three findings invisible to standard accuracy: (i) four of nine models achieve $\geq$79% VQA accuracy while exhibiting shortcut rates above 39%, i.e., correct answers whose reasoning the judge marks as unfaithful; (ii) chain quality significantly predicts answer correctness for seven of nine models, but the two exceptions (Claude Sonnet 4.6, InternVL3.5-8B) reveal qualitatively distinct failure modes, terse output vs. verbose-decorative reasoning, that benchmark accuracy alone conflates; (iii) SFT on SpatialChain improves Qwen3-VL-8B by +6.2 pp in-domain and reduces its shortcut rate to 22%, while a stylistic specialization effect on external benchmarks motivates replay-augmented training as mitigation. The faithfulness judge is validated against 198 human-annotated items, where judge-human agreement matches human-human agreement, and against a second judge from a different provider, which preserves the model ranking ($ρ$ = 0.88). Data, generation scripts, and evaluation code are released at https://github.com/spatialchain/SpatialChainBenchmark.

Vision-language models (VLMs) can detect that an object has rotated across views, but cannot reliably tell by how much. We introduce OR-Bench, a fine-grained benchmark for object-rotation reasoning with eight tasks covering rotation detection, rotation magnitude estimation, and multi-view rotation reasoning. Across 12 VLMs, the gap is stark: the strongest models approach 100% accuracy on detection, yet even coarse magnitude estimation is near chance. When asked for exact angles, models place 91.8--100% of their predictions on just $0^\circ$, $90^\circ$, and $180^\circ$, a failure we term canonical-angle collapse. This collapse persists even without visual input. Representation probing shows that missing information is only part of the explanation. Although rotation information becomes less recoverable at finer granularity, substantial coarse-grained information remains, and a simple linear probe outperforms the models' generated answers. This suggests that VLMs underuse rotation information they already encode. We therefore propose RotationCue, a lightweight decoder that recovers coarse rotation information from the VLM's own frozen representations and feeds it back to the model as intermediate textual context. Across three VLMs, RotationCue improves every model--task combination on OR-Bench, raising macro-average accuracy by 7.9--12.6 points while preserving general capabilities.

Visual entailment (VE) asks whether an image supports, contradicts, or leaves undecided a textual hypothesis. Strong results come from fine-tuning large vision-language models on labelled data, while zero-shot and hybrid approaches remain far behind. A VE hypothesis often bundles several visual claims, yet existing zero-shot methods reason over it as a single unit. We propose Atomic Visual Entailment (AVE), which decomposes the hypothesis into atomic facts, produces candidate predictions from both the full hypothesis and its facts using frozen vision-language models, and predicts the final label with a lightweight classifier trained only on how those candidates behave. We find that decomposition helps only when the hypothesis context is preserved: judging facts in isolation is worse than not decomposing at all. Full-hypothesis and atomic prediction make complementary errors, and learning which to trust recovers far more of that complementarity than majority voting, reaching 0.803 test accuracy on SNLI-VE without fine-tuning any vision-language model. AVE also localises the visual evidence behind its prediction without region-level supervision. These results suggest that learning which candidate prediction to trust can close much of the gap to fine-tuned systems, offering a practical alternative where fine-tuning a vision-language model directly would need more labelled data or compute than is available.

We introduce Self-Saliency, a method for training Vision-Language Models (VLMs) to increase the alignment between their visual attention and the image regions mentioned in their reasoning. Self-Saliency uses a grounding model to localize the objects mentioned in each reasoning step and treats the resulting areas as supervision for the model's visual attention. Previous work on steering visual attention determines target image regions based solely on the image and question. In contrast, we show that conditioning the target regions on the model's generated reasoning improves downstream performance. For proper evaluation, we build a unified, broad suite of 25 visual reasoning benchmarks, where we reproduce the results of previous methods. We find that Self-Saliency significantly outperforms both prior attention-steering methods and baselines that ground image-level text, achieving both a better average rank and a better mean score. Post-training analysis shows that the model primarily adapts its reasoning text to existing attention patterns, producing shorter steps that refer to larger regions. Nevertheless, when controlling for generated text, attention to grounded regions increases significantly across the relevant layer. Finally, we identify a consistent geometric bias in VLM visual attention toward the image border. However, our ablations show that Self-Saliency's gains cannot be explained by simply aligning attention with the center of the image, highlighting the importance of aligning visual attention with the regions mentioned in the model's reasoning.

Recent studies show that artificial intelligence (AI) with language and vision capabilities still experiences limitations in spatial reasoning. In this paper, we have studied the spatial capabilities of advanced generative AI to understand the rotations of objects in 3D space, utilizing AI's image processing and language processing features. We trained and examined the spatial intelligence of a generative Agentic AI model (GPT-5.6) to understand the spatial rotation process with rotation diagrams based on the revised Purdue Spatial Visualization Test: Visualization of Rotations (Revised PSVT:R). We improvised the Revised PSVT:R by superimposing additional graphical and contextual features to evaluate how different Chain-of-Thought (CoT) reasoning strategies influence model performance. The results indicate that structured CoT reasoning improves the spatial reasoning performance of the base GPT-5.6 model in both datasets (PSVT:R and PSVT:R with coordinate system). We used three CoT approaches - (1) Structured CoT, (2) few-shot Structured CoT, and Structured CoT with Self-optimized Prompt. The three CoT approaches evaluated in this study showed no significant performance difference. Results showed that combining structured CoT reasoning with relevant contextual information leads to considerable improvements in VLM performance on 3D rotation tasks, demonstrating the potential of agentic AI for more effective spatial reasoning. However, when contextual information is removed, structured CoT reasoning alone provides limited improvement, and the models continue to exhibit notable difficulties in understanding spatial transformations. These findings suggest that effective spatial reasoning in VLMs relies on the integration of visual, textual, and reasoning-based information in future agentic AI systems for spatial intelligence.

Spatial reasoning benchmarks evaluate vision-language models across diverse tasks, but task-level scores do not reveal which underlying capabilities account for success or failure. Each task requires recovering spatial evidence, representing geometry, and reasoning over it. We disentangle these capabilities by comparing predicted and ground-truth spatial context under a shared schema and coordinate contract. This comparison reveals four recurring sources of error: inaccurate perception, missing information in the spatial context, selection of the wrong measurement, and errors in reference frames or in tracking position and orientation. Guided by this diagnosis, we develop CROSS, a training-free library of typed geometric operators and spatial skills that function over available evidence to support reliable video spatial reasoning. The resulting library supplies verified context to non-coding VLMs or callable skills to a SpatialClaw agent. We evaluate \methodname{} on five benchmarks. \methodname{} raises the average score from 55.9% to 60.2% on ReVSI and improves the SpatialClaw result from 62.8% to 66.3% on DSI-Bench. These gains demonstrate that explicit handling of spatial conventions can repair systematic reasoning failures without additional training.

Reliable visual reasoning requires composing multiple visual observations and returning consistent answers to logically equivalent questions. We introduce Hob-VL, a benchmark for visually grounded Boolean reasoning. Hob-VL comprises two tasks: (1) evaluating whether a Boolean rule holds in an image, and (2) identifying the (unique) object satisfying a Boolean description. Hob-VL contains 6,000 human-verified balanced Yes/No questions, each defined by a Boolean combination of ten visual statements, across 1,000 generated scenes and 46 diverse labeled photographs, along with 1,000 object-identification questions over the same photographs. Our question families are deliberately constructed to challenge reasoning through misleading local cues and nested logical operations, and include symbolic and structured natural-language presentations. Across eight model configurations with thinking disabled or minimized, Boolean accuracy ranges from 48.52% to 50.57%, while the identification accuracy reaches at most 43.0%. A thinking-enabled GLM configuration achieves uneven gains while retaining substantial errors and inconsistencies. Hob-VL exposes these failures through executable reference answers and matched evaluations.

Despite the strong performance of Vision-Language Models (VLMs) on a wide range of visual question answering (VQA) tasks, these models consistently struggle to understand negation and produce incorrect answers when questions involve negated clauses. To address this limitation, we propose Skeleton-and-Strategy Prompting (\textbf{SSP}), a training-free, in-context learning method that improves VLM negation understanding capabilities without any parameter updates. Given a negation question, our method first abstracts the underlying question structure into a skeleton, retrieves a small set of same-skeleton questions from a lightweight question pool, then prompts the VLM to analyze their shared negation pattern and synthesize a single-sentence answering strategy. The skeleton and strategy are prepended to the test sample to guide the model correctly tackle the negation problems. Experiments on multiple negation VQA benchmarks show that SSP achieves state-of-the-art performance on negation-focused VQA tasks while remaining computationally efficient.

Cardiovascular magnetic resonance (CMR), including cine imaging, is a reference standard for the noninvasive assessment of cardiac morphology and ventricular function. Cine CMR interpretation integrates qualitative visual assessment with quantitative measurements of ventricular volumes, ejection fraction, myocardial mass, wall thickness, and regional wall motion. Current medical vision-language models (VLMs) cannot reliably derive quantitative measurements from multidimensional cine images without analysis tools. We present CineMR, a tool-augmented VLM that invokes cardiac image-analysis tools and integrates their outputs into interleaved reasoning for quantitative CMR assessment. We also construct a multi-cohort visual question answering benchmark covering quantitative metric extraction, multiclass diagnosis, and differential diagnosis, together with tools for segmentation, phase selection, volumetry, morphometry, and regional wall motion analysis. CineMR is trained with supervised fine-tuning (SFT) on tool-interaction traces followed by Group Relative Policy Optimization (GRPO) with conditional tool-use rewards. On the multi-cohort cine CMR benchmark, CineMR achieves 35.9% pass@1 and 58.9% pass@4, compared with 1.5% pass@1 for the Qwen3-VL-8B backbone and 0.0% and 7.0% pass@1 for LLaVA-Med v1.5 and MedGemma-4B, respectively. Correct tool invocation reaches 99.8% after GRPO, up from 78.9% after SFT. Live tool outputs improve ventricular measurement accuracy by 20.4--23.7% over direct model predictions, and removing all tools reduces pass@1 from 35.9% to 27.9%. These results highlight the importance of reliable tool use for quantitative cine CMR reasoning and support CineMR as a promising approach for assistive cardiac image assessment. Code, benchmark resources, and model weights are available at https://github.com/AI-MIND-Lab/CineMR.

An answer candidate in a masked diffusion MLLM can stabilize while its rationale is still unfolding. We distinguish retrospective stabilization of the logged candidate from token commitment, and examine these two clocks relative to rationale generation. Analyzing our results across three visual question-answering benchmarks, we find that 89.4-98.1% of the rationale-side canvas remains unwritten at stabilization in single-block, EOS-suppressed LaViDa runs. On VBench, reducing block length from 128 to 8 changes this fraction from 89.4% to 1.7%, together with answer coverage and the eligible observation window. Under EOS-enabled prompting, direct instructions improve Nemotron's overall accuracy by 15.0 and 19.5 percentage points on M3CoT and ScienceQA, but reduce LaViDa/VBench accuracy by 11.0 points. A symmetric decomposition associates the larger absolute component of each change with coverage rather than conditional accuracy. Matched-canvas image ablations measure visual sensitivity alongside answer stabilization, separating the two temporal readouts. Together, these measurements distinguish answer stabilization, rationale unfolding, and visual sensitivity, and identify coverage as the larger component of the prompting differences.

Vision-Language Models (VLMs) have shown strong multimodal reasoning capabilities, yet whether they truly capture the physical consistency underlying real-world dynamics remains unclear. Existing benchmark paradigms often suffer from fragmented evaluation, focusing on isolated cognitive stages while overlooking the inherent synergy between perception, reasoning, and physical judgment. The lack of a holistic perspective limits the ability to diagnose whether VLMs can reliably evaluate the physical authenticity of emerging generative models. To address these issues, we introduce PhysVista, a benchmark designed to evaluate physical intelligence in VLMs through a closed cognitive loop framework inspired by the human seeing-reasoning-assessment process. PhysVista restores this loop by jointly evaluating physical state perception, physical dynamics reasoning, and physical plausibility assessment. It further distinguishes event-level reasoning and scale-level reasoning to enable fine-grained analysis of physical understanding. In addition, PhysVista incorporates both real-world and AI-generated videos, allowing evaluation across diverse domains and emerging generative scenarios. Extensive experiments across a diverse set of VLMs reveal substantial limitations in physical reasoning and plausibility assessment, highlighting a persistent gap between visual recognition and genuine physical understanding, and pointing toward more principled designs for physically grounded multimodal intelligence.

Robot navigation in dynamic, human-centered environments requires socially-compliant decisions grounded in robust scene understanding. Recent Vision-Language Models (VLMs) exhibit promising capabilities such as object recognition, common-sense reasoning, and contextual understanding, capabilities that align with the nuanced requirements of social robot navigation. However, it remains unclear whether VLMs can accurately understand complex social navigation scenes (e.g., inferring the spatial-temporal relations among agents and human intentions), which is essential for safe and socially compliant robot navigation. While some recent works have explored the use of VLMs in social robot navigation, no existing work systematically evaluates their ability to meet these necessary conditions. In this paper, we introduce the Social Navigation Scene Understanding Benchmark (SocialNav-SUB), a Visual Question Answering (VQA) dataset and benchmark designed to evaluate VLMs for scene understanding in real-world social robot navigation scenarios. SocialNav-SUB provides a unified framework for evaluating VLMs against human and rule-based baselines across VQA tasks requiring spatial, spatiotemporal, and social reasoning in social robot navigation. Through experiments with state-of-the-art VLMs, we find that while the best-performing VLM achieves an encouraging probability of agreeing with human answers, it still underperforms simpler rule-based approach and human consensus baselines, indicating critical gaps in social scene understanding of current VLMs. Our benchmark sets the stage for further research on foundation models for social robot navigation, offering a framework to explore how VLMs can be tailored to meet real-world social robot navigation needs. An overview of this paper along with the code and data can be found at https://larg.github.io/stars.

We push the frontier of large-scale spatial intelligence in Vision-Language Models (VLMs) and introduce the first benchmark that probes geographical layout understanding from real-world videos, spanning up to 1km distances. Inspired by the cognitive science literature, we evaluate models against the hierarchical stages of human spatial awareness: anchoring via landmarks, connecting them through routes, and integrating these into global mental maps. Extensive experiments reveal a fundamental divergence in how current AI models process spatial information. Instead of utilising true path integration or forming geometric survey knowledge, we find that VLMs rely almost entirely on 2D visual recognition and text-matching to bypass complex spatial reasoning. The benchmark is publicly available at https://perception-test-challenge.github.io/kilometervision.html.

Spatial intelligence requires maintaining a coherent understanding of the world as the embodied agent moves. Like humans, the agent must use its own motion to interpret changes across observations and update object locations and spatial relations accordingly. Despite spatial post-training having substantially broadened the spatial intelligence of vision-language models (VLMs), they still struggle with two atomic spatial capabilities: tracking self-motion and mapping the surrounding world during motion. To address this gap, we provide dense multi-turn supervision over interleaved atomic capabilities within each training episode, mimicking the visual experience of a continuously moving agent that reasons as it observes. To scale this up, we synthesize 11,738 motif-driven camera trajectories over a broad range of 3D scenes, supporting self-motion tracking, persistent object mapping, and rich spatial operations within each visual experience. By training models to reason over these atomic questions, our URUQI$_{\mathrm{Syn}}$-8B improves accuracy from 15.84% to 47.73% on our Uruqi benchmark comprising 52k questions across 2.7k episodes. URUQI-SI-Mix-8B further reaches 50.41%, comparable to the 50.08% achieved by GPT-6 Astra. Trained solely on our synthesized data, URUQI$_{\mathrm{Syn}}$-8B achieves an average relative accuracy improvement of 17.13% over its InternVL3-8B backbone across three external spatial benchmarks. These results highlight continuous visual experience as a scalable source of supervision for developing spatial cognition in VLMs.

This paper studies \textbf{thinking--answer consistency} in vision-language models. We focus on Visual Intention Grounding, where a model infers a target object based on a human intention query and predicts a bounding box. We reveal that previous IoU-based reinforcement learning (RL) frameworks suffer from ``thinking drift'', where the model produces a correct bounding box, despite having an incorrect reasoning process pointing to a different target object. Thus, we propose \textbf{Rita} (\textit{ReInforcing Thinking--Answer consistency}) as a novel RL paradigm to tame the drift. Specifically, Rita introduces two reasoning-label-free RL rewards, constructed from the conditional probability of reference answers: a \textbf{thinking reward} and a \textbf{consistency reward}. It also adopts a difficulty-aware \textbf{data filtering} strategy that selects informative easy-to-medium samples for RL using rollout error rate and reward variance. Extensive experiments on EgoIntention and the new RefEgo-Int benchmarks show that Rita performs consistently superior to the supervised finetuning approaches and vanilla RL-finetuned frameworks.

Can structured text replace vision for diagram reasoning? A wrong answer after textualization can arise because the representation omits information the question needs, or because the solver fails to use information that is present. We introduce a diagnostic protocol to distinguish these explanations. Using the same solver model and generation settings, we compare three input conditions: the original image, question-blind structure extracted by a vision-language model, or gold structure derived from the diagram source. Validity-triggered recovery tests truncation and schema failure, question-relevant fidelity measures preservation of answer-critical structure, and matched edge interventions test the effect of error location. On a reserved holdout of 240 public FlowGen diagrams, evaluated under a frozen protocol, gold structure reaches 87% accuracy while direct vision and learned text both remain below 30%. The aggregate comparison includes source-derived relation labels that may not be printed in the image and uses different learned and gold graph encodings, so it does not isolate extraction error alone. Retrying only invalid extractions makes nearly every public representation schema-valid yet leaves accuracy essentially unchanged. The public learned-text deficit relative to gold more than doubles with structural difficulty. Question-relevant topology predicts correctness better than whole-graph topology. In an exposed intervention study, a single answer-relevant edge edit reduces the primary solver's original-answer accuracy to near zero, while matched irrelevant edits largely preserve it. Supplied structure requires fewer solving tokens than vision, but learned acquisition removes this advantage at single use. These comparisons motivate evaluating acquired text by the answer-relevant evidence it preserves and by the solver's ability to use that representation.