ECCV2026 VLM Reasoning AI paper notes paper summaries Reasoning Multimodal/VLM Reinforcement Learning Segmentation Layout & Composition Agents

🧠 VLM Reasoning¶

🎞️ ECCV2026 · 164 paper notes

📌 Same area in other venues: 📷 CVPR2026 (150) · 🔬 ICLR2026 (112) · 💬 ACL2026 (32) · 🧪 ICML2026 (31) · 🤖 AAAI2026 (10) · 🧠 NeurIPS2025 (30)

🔥 Top topics: Reasoning ×143 · Multimodal/VLM ×35 · Reinforcement Learning ×8 · Segmentation ×7 · Layout & Composition ×7

Video-Holmes: Can MLLM Think like Holmes for Complex Video Reasoning?: Video-Holmes tests whether models can independently locate and connect scattered clues through 1,837 questions about 270 suspense short films; Gemini-3.0-Pro leads the main table at 49.6%, compared with 96.3% for humans watching each film once.
Chain-of-Visual-Thought: Teaching VLMs to See and Think Better with Continuous Visual Tokens: CoVT teaches a VLM to generate continuous visual tokens supervised by segmentation, depth, and semantic experts before answering, replacing purely verbal intermediate observations with 16 latent perceptual states and raising Qwen2.5-VL-7B from 74.5 to 80.0 on CV-Bench, including a 14.0-percentage-point gain on its depth subtask.
LEGO-Puzzles: How Good Are MLLMs at Multi-Step Spatial Reasoning?: LEGO-Puzzles progressively evaluates elementary spatial questions, assembly planning with distractors, and multi-turn intermediate-state generation, revealing substantial gaps across 29 multimodal models: GPT-5 trails humans by 21.8 percentage points on the Lite subset, while its planning accuracy falls from 90% at one step to 0% at eight steps.
TIR-Bench: A Comprehensive Benchmark for Agentic Thinking-with-Images Reasoning: TIR-Bench expands visual reasoning evaluation beyond crop-based search to 13 image-operation tasks with 1215 examples, showing substantial benefits from tool execution while the strongest o3-TU configuration still scores only 46.0 overall and struggles with complex spatial planning and tool selection.
Vision-as-Inverse-Graphics Agent via Interleaved Multimodal Reasoning: VIGA repeatedly writes code, renders, actively inspects, and edits, using semantic skills and bounded historical memory to maintain editable scenes without training, and reports a 124.70% average improvement over one-shot generation on BlenderBench.
Reinforcing Video Reasoning with Focused Thinking: TW-GRPO reallocates learning strength using within-group token distributional differences at aligned positions and grants partial credit to multi-answer predictions under a strict false-positive constraint, improving Qwen2.5-VL-7B's CLEVRER accuracy from 41.1% with matched GRPO to 50.4% using only 1K CLEVRER training samples, although gains do not extend to every video benchmark.
UniREditBench: A Unified Reasoning-based Image Editing Benchmark: UniREditBench extends image editing evaluation to real-world interactions and game-rule reasoning, constrains the evaluator with textual and target-image references, and trains UniREdit-Bagel with synthetic chain-of-thought data to achieve an in-domain overall score of 78.15, while remaining behind the strongest closed-source models on out-of-distribution RISEBench.
Embed-RL: Reinforcement Learning for Reasoning-Driven Multimodal Embeddings: Embed-RL uses retrieval feedback from a frozen embedder to train a separate reasoner and supplies the embedder with image boxes, video keyframes, and text keywords from its chain-of-thought, allowing Embed-RL-4B to reach 68.1 overall on MMEB-V2, 3.6 points above UME-R1-7B.
Video Streaming Thinking: VideoLLMs Can Watch and Think Simultaneously: VST moves video reasoning ahead of the user query, continually updating textual memory through causally trained streaming thoughts, enabling VST-7B to reach 79.5% on StreamingBench and 41.9% accuracy with 0.56-second query-response latency on VideoHolmes.
LaViT: Aligning Latent Visual Thoughts for Multi-modal Reasoning: LaViT makes Qwen2.5-VL-3B generate four continuous visual tokens before answering, distills visual semantics and attended regions from a 32B teacher, and gradually opens the direct image-access path to prevent bypassing those tokens, improving BLINK Relative Depth from 61.29% to 78.23%.
Towards Spatial Trace with Reasoning in Vision-Language Models for Robotics: This work proposes RoboTracer, a 3D-aware VLM that learns "3D spatial referring + metric measurement" via a regression-supervised scale decoder and a universal spatial encoder with plug-and-play geometric inputs. It is then elevated to multi-step, metric-grounded spatial trace reasoning using GRPO reinforcement fine-tuning with metric-sensitive process rewards. RoboTracer outperforms Gemini-2.5-Pro by 36 percentage points on the self-built TraceSpatial-Bench and can directly interface with motion planning to drive real robots such as the UR5 and the G1 humanoid.
RePlan: Reasoning-Guided Region Planning for Complex Instruction-Based Image Editing: RePlan uses a vision-language model to determine which regions to edit and how, then executes the edits jointly through attention region injection without retraining the editor, improving Flux.2-Klein-9B Overall from 3.45 to 3.92 on IV-Edit, primarily through better regional consistency.
Focusing by Contrastive Attention: Enhancing VLMs' Visual Reasoning: CARVE extracts task-relevant regions through the ratio between the same model's attention to a specific question and a general image description, then masks, crops, magnifies, and answers again, raising LLaVA-1.5-7B's V* accuracy from 38.7% to 66.5% without training but requiring three generation passes.
EgoVITA: Learning to Plan and Verify for Egocentric Video Reasoning: EgoVITA explicitly decomposes egocentric video reasoning into a two-stage "ego-planning + exo-verification" process. Trained via GRPO reinforcement learning coupled with two dense reward signals (ACMG predictive cross-modal alignment + confidence verification preference optimization), it outperforms Qwen2.5-VL-7B by +7.7 points on EgoBlind using only 52k samples, while maintaining zero degradation in exocentric video understanding performance.
Thinking in Streaming Video: ThinkStream makes a video model watch, reason, and decide whether to respond chunk by chunk, retains historical reasoning as a semantic substitute for evicted visual tokens, and jointly optimizes answers and response timing through streaming RLVR, improving Qwen2.5-VL-3B's overall OVO-Bench score from 51.00 to 59.66.
SyncLoop: A Multimodal Dual-Loop Framework for Self-Improving Mathematical Reasoning: SyncLoop jointly generates more complex mathematical images and questions, then selects training examples using the current model's answer error rate, improving Qwen2-VL-7B Geo-Sub-Aux accuracy from 40.4% to 52.4% after three evolution rounds (Table 2, p. 10).
EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models: EndoCoT recursively updates continuous thought states in a diffusion model's multimodal encoder, using terminal textual supervision and staged visual supervision to coordinate reasoning with generation; under task-specific training, it reports 92.1% average accuracy, 8.3 percentage points above DiffThinker, without simply increasing denoising steps.
GRADE: Benchmarking Discipline-Informed Reasoning in Image Editing: GRADE turns image editing into a joint test of disciplinary knowledge and diagram constraints, evaluating 20 models on 520 samples with three scoring dimensions; the best strict accuracy is only 46.2%, while explicit editing instructions substantially reduce but do not eliminate failures.
WeatherReasonSeg: A Benchmark for Weather-Aware Reasoning Segmentation in Visual Language Models: WeatherReasonSeg combines controlled synthetic weather with five-dimensional reasoning queries for real adverse-weather images into 44,721 image-query pairs, revealing failures in both mask prediction and language-conditioned target grounding, with the best evaluated reasoning model reaching only 28.9 gIoU on real nighttime scenes.
Reliable Reasoning in SVG-LLMs via Multi-Task Multi-Reward Reinforcement Learning: CTRL-S trains Qwen3-VL-8B to plan drawing steps before generating corresponding SVG groups and jointly learns text-conditioned generation, image reconstruction, and code refinement, improving SArena-Icon Image-to-SVG DINO from 0.903 after SFT to 0.980 and success rate from 90.12% to 99.93%.
SNOW: Spatio-Temporal Scene Understanding with World Knowledge for Open-World Embodied Reasoning: SNOW uses point-cloud-guided segmentation to combine object patches with 3D shape, position, and time as STEP tokens, enabling an existing VLM to query a persistent 4D scene graph without additional training and raising Gemma3-4B-IT's VLM4D overall score from 39.9% to 72.6%.
VisReason: A Large-Scale Dataset for Visual Chain-of-Thought Reasoning: VisReason extends visual question answering into multi-round supervision that describes a scene, localizes evidence, crops for inspection, and then answers, providing 489K examples including a 165K VisReason-Pro subset and improving Qwen2.5-VL-7B's sample-size-weighted judge score on the Visual-CoT evaluation suite from 0.770 to 0.802, with gains concentrated in fine-grained recognition and spatial relations.
Zoom-IQA: Image Quality Assessment with Reliable Region-Aware Reasoning: Zoom-IQA equips a no-reference image quality assessment model with selective cropping and re-inspection, combining GR-IQA supervision with stabilized reinforcement learning to improve explanation quality under model-based evaluation while retaining competitive score correlations, rather than claiming the best scores on every benchmark.
A4-Agent: An Agentic Framework for Zero-Shot Affordance Reasoning: A4-Agent assigns interaction imagination, functional-part selection, and pixel-level localization to specialized pretrained models, reaching 70.52 gIoU on ReasonAff without affordance-specific fine-tuning, at approximately 32 seconds per inference for the full pipeline.
HyLaR: Hybrid Latent Reasoning with Decoupled Policy Optimization: HyLaR achieves hybrid discrete-continuous reasoning by introducing control tokens to seamlessly switch between discrete text tokens and continuous visual latent vectors. For efficient reinforcement learning in hybrid action spaces, it proposes DePO (Decoupled Policy Optimization), which utilizes von Mises-Fisher (vMF) spherical modeling and decoupled trust-region clipping to address the geometric and variance mismatch issues of standard RL, significantly outperforming existing methods on fine-grained perception and general multimodal understanding benchmarks.
PosterCopilot: Toward Layout Reasoning and Controllable Editing for Professional Graphic Design: PosterCopilot trains a 7B vision-language model to output layout JSON through perturbed supervision, geometric reinforcement learning, and aesthetic reinforcement learning, then enables professional poster iteration through a generative agent restricted to target layers, reaching 54.31% IoU on zero-shot Crello evaluation.
Reasoning Path and Latent State Analysis for Multi-view Visual Spatial Reasoning: A Cognitive Science Perspective: The paper introduces ReMindView-Bench and combines scene-graph matching of reasoning traces, self-consistency checks, representation probing, and token entropy analysis to show that VLMs struggle particularly with cross-view relational integration and subsequent inference, not merely with recognizing objects in individual images.
3D-Layout-R1: Structured Reasoning for Language-Instructed Spatial Editing: 3D-Layout-R1 trains models to update JSON 3D scene graphs while reasoning, then refines their outputs with IoU, collision, and format rewards through GRPO, raising Qwen3-8B's mean IoU on perfect-input sorting from 0.850 with Vanilla GRPO to 0.924, although gains and metric trade-offs depend substantially on the task.
BrepCoder: A Unified Multimodal Large Language Model for Multi-task B-rep Reasoning: BrepCoder aligns B-rep surface geometry and topology with Python-like CAD programs, learns shape-to-construction correspondences through reverse engineering, and adapts to completion, correction, and QA, reducing median Chamfer distance from CADCL's 0.972 to 0.464 on DeepCAD reverse engineering, both on the \(10^{-3}\) scale, without leading every task or metric.
GeoBrowse: A Geolocation Benchmark for Agentic Tool Use with Expert-Annotated Reasoning Traces: GeoBrowse combines weak-cue visual geolocation and entity-obfuscated multi-hop questions in a 300-instance benchmark, using expert evidence traces to diagnose the GATE tool workflow; Gemini-3-Pro reaches reported aggregate accuracies of 48.2% and 34.7% on the two levels in Table 4.
HART: High-Resolution Annotation-Free Reasoning Technique through a Closed-loop Framework: HART first predicts key regions on a downsampled full image, then deliberately withholds the full image and lets the model answer using only those regions cropped from the original high-resolution image, so that "is the answer correct" becomes a self-verification signal for "is the localization accurate"; an advantage-weighted AP-GRPO then optimizes localization directly without any bounding-box annotations, lifting Qwen2.5-VL-7B from 42.3% to 62.4% on MME-RealWorld-Lite.
Make Geometry Matter for Spatial Reasoning: GeoSR weakens 2D appearance shortcuts during training and allocates geometry contributions per token and channel during fusion, making an existing geometry branch useful for spatial question answering and achieving official average scores of 51.9 on VSI-Bench and 66.1 on DSR-Bench.
PuzLM: Solving Jigsaw Puzzles with Sequence-to-Sequence Language Models: PuzLM turns image pieces into symbolic sequences through lightweight border tokenization, then uses a standard encoder-decoder model to predict each piece's grid position, achieving 92.2% absolute accuracy and 87.1% perfect accuracy on ImageNet 3x3.
REVEAL: Reasoning-Enhanced Forensic Evidence Analysis for Explainable AI-Generated Image Detection: REVEAL first uses eight lightweight expert models to extract structured forensic evidence, has a large model consolidate it into chain-of-evidence (CoE) annotations, and distills a 60K REVEAL-Bench; it then trains an MLLM with a two-stage recipe (CoE tuning followed by R-GRPO with an evidence-grounded composite reward), reaching 95.31% in-domain and 94.96% cross-domain on GenImage while producing explanations that can be checked against low-level evidence.
The Cost of Reasoning: Chain-of-Thought Induces Overconfidence in Vision-Language Models: Across four multimodal benchmarks and four open-weight VLMs, this paper shows that chain-of-thought prompting and reasoning-trained models improve task accuracy while substantially degrading the ranking quality of uncertainty estimates that rely on answer-token likelihood (ATL); it attributes this to implicit answer conditioning — the reasoning trace commits to a conclusion before the answer is generated, inflating answer-token likelihood regardless of correctness — whereas agreement-based consistency stays robust and often improves under reasoning.
Think While Watching: Online Streaming Segment-Level Memory for Multi-Turn Video Reasoning in Multimodal Large Language Models: Think While Watching (TWW) trains a multimodal large language model to write reusable text memory for each video segment, then combines streaming causal masks, independent positions, and dual KV caches to support continuous watching and multi-turn answers; Qwen3-VL-4B reaches 60.04% online single-turn and 57.40% multi-turn accuracy on StreamingBench, with the latter using 56.10% fewer output tokens than offline Thinking but scoring 1.12 percentage points lower.
VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations: VIEW2SPACE generates cross-view questions and geometrically verifiable reasoning traces from controlled 3D scenes, training Qwen3-VL-4B to output intermediate object boxes and reach 64.93% MCQ accuracy and 69.34% grounding mIoU, while increased training scale still struggles with difficult compositional reasoning.
VisionCoach: Reinforcing Grounded Video Reasoning via Visual-Perception Prompting: VisionCoach turns visual prompting from an inference-time tool into a training-time coach: a VP-Selector picks one prompt type (red circle / darkening / frame numbering, etc.) for hard samples only, overlays it on key frames and re-samples trajectories, and self-distillation internalizes those high-reward trajectories into the policy, while an object-aware spatial grounding reward keeps object identity consistent — so at inference the model takes a single forward pass over the raw video, improving +15.0 mAM / +25.1 mLGM over Qwen2.5-VL-7B on V-STAR and surpassing GPT-4o.
AnE: Pushing the Reasoning Frontier of Multimodal LLMs via Anchor Evolution: AnE retrieves real problems around a student's persistent failures, uses teacher hints to elicit the student's own successful reasoning trajectories, and removes external assistance through hint-free reinforcement learning, raising Qwen2.5-VL-7B's average accuracy across eight benchmarks from 51.6% to 61.9% after three evolution rounds.
Beyond Where to Look: Trajectory-Guided Reinforcement Learning for Multimodal RLVR: TGRL mixes a small number of complete expert reasoning trajectories into the student's RLVR rollout groups, using off-policy probability-ratio correction, adaptive token reweighting, and successful-trajectory filtering to connect observed facts to reasoning, improving Qwen2.5-VL-7B's five-benchmark DAPO average from 58.15 to 60.68 in the 39K-data setting.
E-TTS: A New Embodied Test-Time Scaling Framework for Robotic Manipulation: E-TTS is a plug-and-play embodied test-time scaling framework that performs joint sampling, history-aware verification, and feedback-guided iterative refinement of the VLA model's reasoning process and actions during inference. Without extra training or data, it increases the success rates of various base models by up to 33.14% in both simulation and real-world scenarios.
GAIA: A Data Flywheel System for Training GUI Test-Time Scaling Critic Models: This paper proposes the GAIA data flywheel system, which collects positive and negative action samples by running real GUI agents on tasks to train an intuitive (non-reasoning) critic model (ICM). During testing, ICM selects the correct action with the highest probability from multiple candidate actions sampled by the agent using a Best-of-N approach. These selected actions are then used to expand the data flywheel to train a stronger ICM-r2, establishing a self-evolving loop.
How Far Are Video Models from True Multimodal Reasoning?: The paper operationalizes "true multimodal reasoning" as a set of zero-shot video generation tasks whose contexts supply only raw material and never the conclusion, builds CLVG-Bench with 6 categories, 47 subcategories and roughly 1,390 entries over four input modalities, and proposes AVE, a judge whose prompt is optimized through semantic set matching; it finds that the best video models clear 60% pass rate on conventional editing tasks but collapse to around 21% once they must supply the physical or logical causality themselves, with multi-turn interaction failing entirely at the second turn.
MultiHaystack: Benchmarking Multimodal Retrieval and Reasoning over 40K Images, Videos, and Documents: MultiHaystack separates finding evidence from answering with it using 46,260 heterogeneous candidates and 747 unique-evidence questions, showing that GPT-5 reaches only 51.41% overall answer accuracy with E5-V cross-modality Top-5 retrieval, so reasoning with supplied gold evidence cannot stand in for end-to-end reliability.
OmniSch: A Multimodal PCB Schematic Benchmark For Structured Diagram Visual Reasoning: OmniSch uses source-driven EAGLE annotation to build a structured reasoning benchmark of 1,854 schematics, showing that better component recognition does not reliably recover correct netlists and that tool-assisted localization mainly improves fine-grained attribute reading.
PerceptionComp: A Video Benchmark for Complex Perception-Centric Reasoning: PerceptionComp builds 1,114 five-way multiple-choice video questions (279 high-scene-complexity real-world videos, 100% manual annotation) by making several perceptual subconditions jointly necessary, so that no single moment or single viewing suffices; humans reach 100% with unrestricted rewatching but drop to 18.97% in a single-view setting, while the best MLLM (Gemini-3-Flash) reaches only 45.96%.
VideoTIR: Accurate Understanding for Long Videos with Efficient Tool-Integrated Reasoning: VideoTIR answers long-video questions inside a multi-turn loop: it starts from a low-resolution, low-fps view, lets a textual router decide whether the current evidence suffices, and otherwise calls hierarchical internal visual tools (a global browsing tool plus a segment→frame→zoom-in grounding chain) whose output is folded back into context; TAGPO then decays the trajectory-level reward onto each individual tool call according to its distance from the final answer, curbing both tool overuse and misuse and improving accuracy and efficiency (1.42 tool calls, 4.6s latency) on Video-MME / MVBench / LongVideoBench.
When Thinking Hurts: Mitigating Visual Forgetting in Video Reasoning via Frame Repetition: FrameRepeat trains a 3.7M-parameter scorer using the correct-answer log-probability gain from repeating one frame, then duplicates the frames most worth reinforcing in place, improving video question answering on Qwen2.5-VL and Qwen3-VL without fine-tuning the backbone or adding reasoning rounds.
Benchmarking Dynamic Affective Reasoning: A Viewer-Centric Video Emotion Dataset: DAR redefines video emotion understanding through event boundaries, 27 viewer-emotion categories, and contextual causal explanations, while DAR-R1, obtained by applying SFT and GRPO to Qwen2.5-VL-3B, achieves 41.5% segment-count accuracy, 52.3% temporal mIoU, and 28.6% emotion accuracy on its test set.
Dense Reward for Multi-View 3D Reasoning with Global Maps and Local Views: To address the persistent issues of MLLMs in multi-view 3D visual question answering—namely, sparse supervision using only the correctness of final answers, inconsistent cross-view reasoning, and erratic view selection jumping—this paper explicitly decomposes the task into three stages: "building a global bird's-eye-view cognitive map \(\rightarrow\) planning the view sequence based on the question \(\rightarrow\) answering from an egocentric perspective." By utilizing geometrically consistent pseudo-maps generated by a frozen 3D Vision Foundation Model (VGGT+SAM3) as global rewards and reference trajectories derived from benchmark metadata as local rewards, trajectory-level optimization is performed under GRPO, driving a 3B model's performance on MindCube from 37.8 to 66.5.
EXPLORE-Bench: Egocentric Scene Prediction with Long-Horizon Reasoning: The paper formalizes a new task, egocentric scene prediction with long-horizon reasoning — given an initial-scene image and a long sequence of atomic actions, a model must imagine and describe the final scene after all actions are executed. It builds EXPLORE-Bench (1,157 instances, action sequences averaging 113 steps and reaching 694, with structured object/attribute/relation annotations) and evaluates 20+ models with a unified object-coverage + attribute-score + relation-score metric, finding that the best model still trails humans by 7.39 points and that naively segmenting the action sequence for stepwise reasoning loses points because models describe only what changed.
Falcon: Functional Assembly and Language for Compositional Reasoning in X-ray: The paper redefines "threat" in X-ray baggage screening from "an object of some category was detected" to "spatially separated components can, by functional compatibility, be assembled into a usable device"; it contributes Falcon-X (a benchmark of ~7,000 real scans plus counterfactual extensions, with dense masks and structured supervision over component presence, functional links, and scene risk) and Falcon, which compresses a variable number of component instances into a fixed structured safety state via segmentation-aware region encoding and three component queries, then injects that state as tokens into an LLM, surpassing the next best method by 30.38 points in cIoU on referring functional grounding (RFG) with 50.45 cIoU.
Foveated Reasoning: Stateful, Action-based Visual Focusing for Vision-Language Models: FoveateR lets a vision-language model trigger continuous crop actions and incorporate high-resolution evidence within one autoregressive trajectory, improving multiple visual question-answering results with fewer visual tokens under low-resolution inputs without requiring human-annotated foveation trajectories.
How Far Are Vision-Language Models from Constructing the Real World? A Benchmark for Physical Generative Reasoning: DreamHouse asks vision-language models to generate executable construction code from multiple views of a house and evaluates it with ten structural tests plus visual similarity, finding a joint pass rate of only 7.1% even in the best configuration.
MultihopSpatial: Multi-hop Compositional Spatial Reasoning Benchmark for Vision-Language Model: The paper introduces MultihopSpatial, a VLM benchmark that chains Attribute / Position / Relation spatial cues into 1- to 3-hop multiple-choice questions (4,500 items, every one with a human-annotated ground-truth bounding box, split evenly across ego-centric and exo-centric views), together with the Acc@50IoU metric (correct answer and predicted box IoU ≥ 0.5) and a 6,791-sample training corpus; across 37 state-of-the-art VLMs the best model reaches only 40.6% Acc@50IoU, while GRPO post-training on the corpus improves both intrinsic spatial reasoning and downstream VLA manipulation success.
NarrativeTrack: Evaluating Entity-Centric Reasoning for Narrative Understanding: This paper introduces NARRATIVETRACK, the first benchmark that evaluates narrative understanding in multimodal large language models from an entity-centric angle: a fully automated pipeline (ensemble detection → ReID identity tracking → highlighted contextual recognition) extracts timestamped human entity representations from raw video, then a Compositional Reasoning Progression (CRP) over "entity existence → entity changes → entity ambiguity" programmatically generates 1,006 questions scored by exact match; the strongest model, Gemini-2.5-Pro, reaches only 83.80% while every open-source model stays below 57%, exposing a trade-off between perceptual grounding and temporal integration.
OmniCoT: A Benchmark for Global and Multi-Step Panoramic Reasoning: OmniCoT turns panoramic question answering into multi-step See-Locate-Move reasoning that requires cross-view evidence, evaluates both answers and reasoning traces, and raises Qwen2.5-VL-7B accuracy on OmniCoT-B from 23.53% to 59.54% through structured chain-of-thought supervision and GRPO.
OmniMapBench: Benchmarking Visual-Centric Reasoning on Diverse Map Documents: OmniMapBench tests visually grounded reasoning with 2,096 manually verified questions across nine map categories and controlled text-description comparisons; the best of 25 evaluated models achieves a final score of 75.03%, while the benchmark retains a Visual Dependency Index of 0.338 under a 1,024-token description budget.
On Test-Time Scaling for Vision-Language Models: This is the first empirical work to systematically investigate whether test-time scaling methods developed for LLMs can be directly transferred to Large Vision-Language Models (LVLMs). Spanning 13 models, 9 scaling strategies, and 6 benchmarks, it overturns the conventional wisdom that "small models do not benefit." Instead, small but strong instruction-following models benefit the most (up to ~+30%, matching or even exceeding larger models). It also reveals two key phenomena: "excessive computation harms perception tasks" and "visual information is encoded into text tokens early in the reasoning chain."
ProLaViT: Learning Progressive Latent Visual Thoughts in Structured Latent Space: ProLaViT distills programmatically synthesized visual-operation trajectories into four continuous latent states and constrains their differences according to step distance, enabling progressive reasoning in Qwen2.5-VL-7B-Instruct without inference-time visual tools and raising BLINK-Jigsaw accuracy from the base model's 59.33% to 76.00%.
SENTRY: SAM2-Enhanced Neighbor-Aware and Temporally Reasoned Memory for Visual Tracking: SENTRY is a training-free, plug-and-play "refine-before-write" module. It validates the temporal plausibility of each candidate mask using short-term cyclic consistency before writing to memory in SAM2-series trackers, replacing the original confidence-based memory update mechanism. When integrated into five strong baselines, it consistently improves performance across nine benchmarks, achieving new zero-shot SOTA on LaSOT/LaSOText/GOT-10k/VOT20/VOT22/DiDi, with the SAM2-L version maintaining a real-time speed of 32.8 FPS on an A100.
STAC: Selective Spatiotemporal Aggregation and Compression for Video Reasoning Segmentation: STAC first feeds independent CLIP frame features through decoupled bidirectional spatial scanning and causal temporal scanning to turn them into spatiotemporally aware representations, then applies adaptive-threshold hierarchical merging in a temporal-then-spatial order to remove roughly 85% of the visual tokens, and propagates segmentation gradients back into the compression decisions via a straight-through estimator — exceeding compression-free baselines on video reasoning segmentation benchmarks such as Ref-DAVIS17 and ReasonVOS at about 1.8× speedup while remaining frame-by-frame streaming compatible.
Syn-GRPO: Self-Evolving Data Synthesis for MLLM Perception Reasoning: Syn-GRPO attaches an online image synthesis service to GRPO training: the MLLM must additionally predict a diversity score and a text description for a new image, a data server keeps the foreground of the original image and re-paints the background from that description to produce replacements whose box labels stay valid, and a diversity reward teaches the model which samples are worth regenerating — lifting both the exploration space and the final accuracy on three visual perception tasks.
Towards Temporal Compositional Reasoning in Long-Form Sports Videos: The paper introduces SportsTime, a long-form sports video benchmark with 14K+ open-ended QA pairs and 50K+ step-wise temporal evidence annotations, together with Chain-of-Time Reasoning (CoTR): training with a temporal-reward GRPO (tr-GRPO) that aligns each reasoning step to a verifiable time anchor, and inference with an anchor-observe-infer loop that retrieves local clips around predicted anchors to verify evidence turn by turn, raising the average accuracy on Qwen3-VL-4B from 25.15% to 29.24% and span IoU from 0.26 to 0.58.
V-REX: Benchmarking Exploratory Visual Reasoning via Chain-of-Questions: V-REX formalizes open-ended multi-step visual exploration as a Chain-of-Questions, turning an unreproducible free-form process score into discrete, checkable judgments by giving each step a finite set of options; on one shared ground-truth chain it separately measures a VLM's Planning (what to ask next) and Following (can it answer along the chain); across 702 samples and 2,504 questions in 4 categories and 15 scenarios, the strongest model (o3) reaches only 86.1 / 84.2 on the two averages, and small models are consistently better at answering than at asking.
VideoSearch-R1: Iterative Video Retrieval and Reasoning via Soft Query Refinement: VideoSearch-R1 proposes an agentic framework that unifies video retrieval and temporal grounding in an iterative interaction loop. It replaces traditional text-level query rewriting with "soft query refinement" in a continuous latent space, and jointly optimizes retrieval and reasoning via GRPO, achieving state-of-the-art performance on three VCMR benchmarks.
ViewFusion: Structured Spatial Thinking Chains for Multi-View Reasoning: ViewFusion splits multi-view spatial reasoning into a two-stage chain-of-thought that aligns views first and answers afterwards (<spatial_thinking> → <thinking> → <answer>), cold-starts it with synthesized reasoning traces via SFT, and then applies GRPO with a composite reward over correctness, format validity and length, raising Qwen3-VL-4B-Instruct from 30.1% to 35.4% on MMSI-Bench (and from 37.0% to 77.0% on MindCube).
What if? Emulative Simulation with World Models for Situated Reasoning: WanderDream is the first large-scale emulative simulation dataset, containing 15.8K panoramic trajectory videos and 158K question-answer pairs, enabling agents to imagine the complete path to target situations through world models without physical movement and answer spatial "what-if" questions.
AdaThinking-E: One-Token Entropy Regulation for Adaptive Thinking: AdaThinking-E optimizes the first token that decides whether to think separately from the subsequent response, using high-entropy exploration followed by low-entropy convergence to reach 26.5 on OCR-Reasoning with its Qwen2.5-based model while matching approximately R-4B's CharXivRQ performance with only 72 output tokens on average.
Before Thinking, Learn to Decide: Proactive Routing for Efficient Visual Reasoning: PRP trains a small vision-language model to predict its own and a larger target model's solvability before lengthy reasoning begins; on MathVista, JRL keeps 60% of requests with the draft and achieves 74.2% accuracy at 26.9 seconds average latency, versus 74.1% and 64.8 seconds for target-only inference, with a reported 2.41x speedup.
Beyond Final Answers: CRYSTAL Benchmark for Transparent Multimodal Reasoning Evaluation: CRYSTAL adds checkable reference reasoning chains to 6,372 visual questions, exposes incomplete reasoning behind correct answers through content coverage and relative-order evaluation, and raises Qwen2.5-VL-3B's Match F1 from 0.480 to 0.633 and accuracy from 39.85% to 47.52% with CPR-Curriculum.
CARE: Causally-Aligned Reasoning Exploration for Medical Large Language Models: Instead of rewarding every medical reasoning trajectory that reaches the correct answer, CARE checks correctness, learnable difficulty, and whether the rationale independently reproduces the answer, then combines online exploration with difficult-experience replay to raise Hulu-Med-7B's average score on PMC-VQA, MedQA, and MMMU-Med from 63.9 to 65.4 and achieve 86.2% expert-rated valid reasoning on a MedQA subset answered correctly by both compared methods.
Cast and Attached Shadow Detection via Iterative Light and Geometry Reasoning: The paper couples three-class shadow segmentation with three-dimensional light-direction estimation through surface-normal-guided feedback, reducing attached-shadow BER from 19.49 for the strongest fine-tuned baseline to 12.92 after three passes, at the cost of richer geometry and class supervision and increased inference time.
Circuit-MLLM: Topological Logic-Guided Latent-Space Visual Reasoning for Circuit Schematic Understanding: Circuit-MLLM organizes training-time circuit knowledge from multiple vision experts into latent-vector supervision ordered by connectivity, enabling Qwen2.5-VL-7B to trace connections without invoking those external experts at inference and raising its average score from 48.28 to 76.20 on a custom benchmark, although exact match on hard connection identification remains only 28.52%.
CL4D: Contrastive Language–4D Pretraining for Vision-Language Reasoning in Dynamic Scenes: CL4D combines per-frame spatial encoding, cross-frame temporal encoding, and language-contrastive learning before connecting its representations to a frozen 7B language model as 4DVLM, achieving 70.32% batch text-to-motion R@1 on HumanOnly, 16.75 percentage points above the strongest reported baseline, while leaving substantial limitations in global retrieval and open-world reasoning evidence.
CoLT: Teaching Multi-Modal Models to Think with Chain of Latent Thoughts: CoLT compresses multimodal textual reasoning into 3 continuous latent states by default using training-time forward text decoding, backward semantic alignment, and adjacent-state prediction, achieving a 79.1 average over eight benchmarks and approximately 10.1x end-to-end speedup over text CoT on MMStar using a single H200.
ConsiSpace: Learning Geometric Consistency Matters for Video Spatial Reasoning: ConsiSpace controls the writing, fusion, and retrieval of video evidence using camera poses and depth, followed by post-SFT training with three types of cross-view consistency rewards (answer, metric, and topology). It achieves scores of 76.6, 53.0, and up to 58.1 on VSI-Bench, OSI-Bench, and MMSI-Video-Bench, respectively.
CoReLIN: Constraint-based Reasoning for Zero-shot Lifelong Interactive Navigation: CoReLIN inverts the traditional role of large language models from low-level action generators to constraint reasoners over structured scene graphs, balancing immediate manipulation cost against long-term navigability to enable zero-shot lifelong interactive navigation in unknown cluttered multi-room environments.
CrossView: Can Vision-Language Models Reason Across Cameras?: To address the pervasive "single-camera assumption" in multimodal evaluation, this paper presents CrossView—the first video QA benchmark spanning autonomous driving, surveillance, ego-exo interaction, and robotics across 2 to 8 concurrent feeds—revealing that leading VLMs exhibit severe performance drops on cross-view spatial-temporal stitching, instance counting, and viewpoint selection.
CVSBench: A Comprehensive Benchmark for Cross-view Spatial Reasoning and Dreaming: Addressing the severe deficiency of current vision-language models in spatial cognition under extreme real-world viewpoint transformations, this paper introduces CVSBench, a benchmark comprising 3,297 image groups and 40,679 QA pairs across satellite-street pairs, demonstrating that explicit 3D visual imagination brings substantially greater spatial reasoning gains than text-only Chain-of-Thought prompting.
Different Changes Require Different Reasoning: Change-Type-Specialized Experts for Robust Change Captioning: Addressing the reasoning confusion caused by unified pipelines treating distinct image changes identically, this paper proposes MEDIC, which routes image pairs through a two-stage gate to specialized key-value memory experts under symmetric/asymmetric consistency and disentanglement constraints, achieving significant gains across standard benchmarks.
Eliciting Self-Verification in Multimodal Reasoning Agents with Reinforcement Learning: Addressing poor tool-calling calibration and severe evidence noise in multimodal reasoning agents, this paper proposes Self-Verification via Reinforcement Learning (SVRL), an RL-only framework that calibrates search timing with weak competence priors, rewards query diversity, and embeds structured self-verification into reasoning traces, eliminating external verifiers at inference time and enabling test-time search compute scaling.
From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models: Addressing the core tension between video generative models' strong visual rendering and fragile physical/symbolic reasoning, this paper introduces VWG-Bench spanning 9 reasoning dimensions and a three-level decoupled evaluation protocol, alongside Vid-PRE—a model-agnostic prompt reasoner optimized via purely text-reward GRPO that substantially boosts downstream reasoning accuracy without architectural modifications.
From Reasoning to Pixels: Grounded Medical Multimodal LLMs for VQA and Segmentation: Addressing the lack of spatial grounding in medical visual question answering, MedREAL introduces Seg-Anchored Reasoning Pooling (SARP) and Reasoning-to-Visual (R2V) fusion to distill clinical multi-step reasoning into SAM mask decoding, achieving 68.49% gIoU and 70.47% cIoU across four imaging modalities on the newly curated MedRAVS-13K benchmark.
Generative Lane Topology Reasoning via Autoregressive Model with Geometry Prior: TopoGPT shifts lane topology reasoning from the discriminative "detect-then-associate" paradigm to an autoregressive sequence generation paradigm. After pre-training the geometric prior of lane graphs on 3.3 million large-scale map scenes, it aligns them with BEV image conditions via a flow-matching perceptual adapter. It comprehensively outperforms prior methods on OpenLane-V2 by significant margins of +6.4 at the lane-level and +11.6 at the point-level, while the generated lane graphs show dramatic improvements in endpoint geometric consistency and structural completeness.
Ground3D-LMM: Fine-Grained 3D Point Grounding and Spatial Reasoning with LMM: Ground3D-LMM introduces a unified point-cloud multimodal architecture and a 2.5M QA dataset that jointly outputs point-level 3D masks and physical metric measurements at both object and part granularity within interactive dialogue.
Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models: This paper introduces Imaginative Perception Tokens (IPT), intermediate visual representations generated by a unified multimodal model that predict unobserved spatial configurations (novel views, side views, or top-down BEV maps) to dramatically enhance 3D spatial reasoning in multimodal language models.
Improving Reasoning in Vision-Language Models via Perception Verified Self-Training: This paper proposes an unsupervised perception-verified self-training framework that disentangles the caption-reasoning-conclusion pipeline, evaluates visual fidelity using FG-CLIP and PaddleOCR via PerceptEval, and employs a two-stage curriculum with caption-guided reasoning enhancement to boost VLM reasoning performance by up to 16% on M3CoT benchmarks.
Information-Regularized Attention for Visual-Centric Reasoning: This paper proposes Information-Regularized Attention (IRA), which applies stochastic regularization based on the information bottleneck principle to visual value states within the attention modules of VLMs. By injecting data-dependent hierarchical noise through a variational inference framework, the model learns to actively control the amount of visual information injected during end-to-end full-parameter fine-tuning, thereby simultaneously mitigating hallucination, weak visual grounding, and catastrophic forgetting.
IQA-T1: Tool-based Visual Evidence Reasoning for Image Quality Assessment: IQA-T1 presents the first tool-based visual evidence reasoning framework for Image Quality Assessment (IQA), equipping MLLMs with a library of 15 specialized perceptual analysis tools and the Q-Tool reasoning dataset to autonomously gather structured visual evidence, substantially alleviating semantic bias and achieving state-of-the-art average performance across seven benchmarks.
Latent Visual Diffusion Reasoning with Monte Carlo Tree Search: The LVDR framework is proposed, which models the cognitive process of gradually converging from uncertainty to certainty in fine-grained skill assessment as latent space diffusion denoising, and extracts interpretable reasoning trajectories using keypoint-guided Monte Carlo Tree Search. It achieves scoring accuracy surpassing SOTA on four athletic and surgical datasets while outputting transparent decision bases.
LatentPilot: Scene-Aware Vision-and-Language Navigation by Dreaming Ahead with Latent Visual Reasoning: LatentPilot internalizes action-conditioned future visual dynamics directly into a VLM decision backbone via two-step future-privileged supervision during training, enabling strictly causal lookahead at inference without external world models and achieving new SOTA performance on R2R-CE, RxR-CE, and real robots.
LineGraph2Road: Structural Graph Reasoning on Line Graphs for Road Network Extraction: Proposing an end-to-end remote sensing road topology extraction framework, LineGraph2Road, which converts a global sparse Euclidean candidate graph into its dual line graph, leverages a Graph Transformer for edge classification, and integrates an explicit overpass segmentation head with coupled NMS to resolve occlusion disconnects and non-planar overpass shortcut artifacts.
Look Less, Think Faster: Joint Token-Compute Adaptation for Multimodal LLMs: SmartVL introduces an end-to-end framework that jointly adapts visual token sequence length alongside LLM compute depth and width, utilizing shared budget encodings and a differentiable FLOPs latency estimator to dynamically co-schedule compute across stages, significantly expanding the accuracy-efficiency Pareto frontier across multimodal benchmarks and diverse compute budgets.
MarineEVT: Advancing Event-Centric Marine Video Understanding via Visual Tool Reasoning: To tackle sparse, unpredictable events and the scarcity of temporal-causal reasoning in underwater videos, this work constructs MarineEVT, the first 20K multi-dimensional event-centric marine video benchmark, and proposes EVT-R1, a multi-turn visual tool-integrated reasoning framework optimized via dual-reward GRPO.
MV-STRIDE: Enabling MLLMs to Master Multi-View Spatial Reasoning via Hierarchical Capability Modeling: Addressing the lack of coherent 3D cognitive maps in current multimodal LLMs under multi-view observations, MV-STRIDE introduces a hierarchical capability modeling framework with cross-view dependency constraints, coupled with cognition question groups and a three-stage training pipeline (SFT-ColdStart-RL), boosting an 8B open-source model to 38.9% SOTA on MMSI-Bench.
On Locality and Length-Generalization in Visual Reasoning: This paper demonstrates that contemporary vision models fail at out-of-distribution length generalization due to shortcut learning encouraged by single-shot global perception, and establishes that strictly local foveated perception coupled with nonlinear recurrence is both necessary and sufficient for robust visual state tracking.
Online Reasoning Video Object Segmentation: Addressing the pitfalls of offline reasoning video segmentation that rely on future frames and global retrospective disambiguation, this paper formalizes the Online Reasoning Video Object Segmentation (ORVOS) task, presents the ORVOSB benchmark with referent-shift annotations, and develops a causal streaming baseline equipped with continually-updating segmentation prompts and a structured temporal token reservoir.
ReasonCLIP-58M: Visually Grounded Commonsense Reasoning Supervision for CLIP: ReasonCLIP-58M presents a non-intrusive two-stage continual pretraining framework powered by 58.9M visually grounded reasoning samples (ReasonLite-42M and ReasonPro-16M), injecting deep commonsense inference capabilities into CLIP without modifying backbone architectures while enhancing both retrieval and compositional reasoning.
Reasoning with Memory: A Temporal Granularity-Adaptive Framework for Training-Free Long Video Understanding: Addressing context window constraints and temporal incoherence in static similarity-based frame sampling, ReMem introduces a training-free dual-level memory-augmented framework that decodes query temporal granularity via LLM priors and dynamically routes keyframes via structural video memory diffusion, establishing new state-of-the-art results across four LongVideoQA benchmarks.
Reasoning-Guided Part-Level Visual Grounding via Reinforcement Learning: To overcome the severe object-level bias in multimodal large language models (MLLMs) where part queries collapse into parent object detections, this paper proposes Object-Part Hierarchical Reflective Grounding (OP-HRG) trained with part-aware GRPO, enabling a 4B model to surpass 7B baselines and SAM3 across cross-dataset zero-shot benchmarks.
RoadBench: Benchmarking MLLMs on Fine-Grained Spatial Understanding and Reasoning under Urban Road Scenarios: RoadBench proposes the first fine-grained spatial understanding and reasoning benchmark in urban road scenarios with road markings as the core, consisting of 8 graded tasks with a total of 3,040 manually verified test cases across dual BEV (satellite) and FPV (car-mounted) perspectives. A systematic evaluation of 20 mainstream MLLMs reveals that existing models severely lack fine-grained spatial understanding capabilities—the best model achieves an F1 score of only 0.355 on BEV lane counting, and even falls short of a rule-based baseline utilizing traffic common sense on lane direction recognition. This exposes a massive capability gap in MLLMs between "seeing" and "reasoning".
RoboTALES: Learning Reasoning-Guided Robot Policies via Task-Aligned Simulated Futures: RoboTALES introduces a single-stage robot learning framework that steers a video diffusion world model via hierarchical sub-goal plans and aligns its latent representations with a frozen VLM critic using differentiable policy optimization (DDPO), directly conditioning an action diffusion policy for robust long-horizon manipulation.
Rule-VLN: Bridging Perception and Compliance via Semantic Reasoning and Geometric Rectification: Addressing the "Goal-driven trap" where embodied navigation agents prioritize physical geometry while ignoring traffic rules, this paper introduces Rule-VLN, the first large-scale urban benchmark enforcing regulatory compliance, alongside the training-free Semantic Navigation Rectification Module (SNRM) that slashes constraint violations by 19.26% and improves task completion by 5.97%.
SafeGuard: A Multi-Agent Perception-Reasoning Framework for Social-Risk AI-Generated Video Detection: To tackle the fundamental "perception-reasoning gap" in detecting high-fidelity social-risk AI-generated videos, SafeGuard coordinates a hierarchical perceptual solver for fine-grained physical forensic cue extraction with a self-reflective verifier for semantic-physical consistency checking, substantially outperforming prior specialized detectors and generalist vision-language models without end-to-end task retraining.
ScAle: Attention Head Scaling as a Minimal Adapter for Spatial Reasoning in Vision–Language Models: ScAle proposes an extremely lightweight adapter for spatial reasoning in VLMs: while keeping the backbone completely frozen, it learns a bounded scalar (constrained within the range of \((1-s_{max}, 1+s_{max})\) via \(\tanh\)) for each attention head and MLP output in every layer. Multiplication scaling is applied only to the activation values of the last token. With only about 1K trainable parameters, ScAle achieves up to a 134.1% relative accuracy improvement on SpatialEval, being more than 2500 times more parameter-efficient than LoRA.
SciIR: A Large-scale Training Dataset and Benchmark for Scientific Image Reasoning Generation: This paper introduces SciIR-82k (80k+ scientific image-text pairs with reverse reasoning chain annotations) and SciIR-Bench (an atomic-checklist-driven fine-grained evaluation benchmark) using Peirce's semiotic trichotomy (Icon/Index/Symbol) as the theoretical framework. By fine-tuning Qwen-Image-SciIR, the final score on SciIR-Bench is improved from 35% to 43%, systematically revealing the severe deficiencies of current T2I models in scientific reasoning.
Sentinel: Embodied Cooperative Spatial Reasoning and Planning: To overcome the failure of embodied multi-agent coordination in large outdoor domains with dynamic threats, this work introduces the Sentinel Challenge benchmark and CoSaR framework, coupling structured spatial memory with VLM-guided route refinement to achieve rapid, hazard-free gathering.
SiPhy: Single-Image Physical Property Reasoning: SiPhy tackles the inability of single-image approaches to perceive 3D volume and material semantics by integrating depth-adaptive pseudo-voxel sampling, VLM-driven part-based contrastive alignment, and heaviness-aware thickness refinement, outperforming multi-view reconstruction methods on mass, density, and Young's modulus prediction without requiring multi-view captures.
SpaR3D-MoE: Adaptive 3D Spatial Reasoning from Sparse Views Meets Geometry-Inductive Mixture-of-Experts: To tackle topological disruption and cross-modal contention in sparse-view 3D reasoning, SpaR3D-MoE introduces adaptive spatiotemporal manifold sampling alongside an instruction-pose-aware heterogeneous Mixture-of-Experts, setting a new state of the art on VSI-Bench, ScanQA, and SQA3D from sparse RGB-only inputs.
SpatialBoost: Enhancing Visual Representation through Language-Guided Reasoning: SpatialBoost converts dense 3D spatial geometry and relations from 2D/multi-view images into hierarchical multi-turn Chain-of-Thought language descriptions (pixel-level → object-level → scene-level), fine-tuning vision encoders through a frozen LLM via dual-channel attention to empower general vision backbones with superior 3D spatial perception without catastrophic forgetting.
Thinking Ahead: Foresight Intelligence in MLLMs and World Model: The authors introduce FSU-QA and its benchmark FSU-Bench to evaluate foresight intelligence in autonomous driving across 9 hierarchical cognitive tasks over a 12-second horizon, establishing a VLM proxy probing paradigm that validates whether world-model-generated futures convey physically and semantically consistent outcomes.
VERDICT: Training-Free Step-Wise Verification of Multimodal Reasoning via Disagreement-Aware Consensus: Addressing error cascade from subtle yet locally plausible hallucinations in multimodal chain-of-thought reasoning, VERDICT introduces a training-free and plug-in step-wise verification framework that casts verification among frozen specialized judges into a coordination game with a closed-form unique Nash equilibrium, leveraging disagreement structure for step filtering and stability ranking to achieve up to a +5.95% accuracy gain across six diverse benchmarks.
Video-HolmesV2: Can MLLMs Reason with Spatio-Temporal Audio-Visual Evidence in Long Videos?: To eliminate visual heuristics and ungrounded speculation in long video benchmarks, this paper introduces Video-HolmesV2 (784 long videos averaging 43 minutes, 4,000 QA pairs) requiring strict spatio-temporal audio-visual evidence grounding, accompanied by an audio-text guided token compression framework that outperforms full-token baselines.
Attention-based Vision-Language Memory for Spatial Reasoning: VLM² aligns semantic features, predicted geometry, and viewpoint information from monocular video before using sliding-window working memory and fixed-capacity episodic memory to retrieve, fuse, and update scene representations, reaching 68.8 on VSI-Bench with video-only inputs, 7.9 score points above VLM-3R.
CabinSI: Omni-Cabin Spatial Reasoning through Explicit Visual Cognitive Maps: CabinSI evaluates spatial relations and referring localization inside and outside vehicles with 3,758 real-world samples, and converts multi-camera observations into query-relevant top-down cognitive maps that raise Qwen2.5-VL-7B overall RelCabin accuracy from 25.31% to 41.17%, while perception errors and cross-cabin grounding remain substantial bottlenecks.
Clue Matters: Empower Video Reasoning with Brain-Inspired Latent Clue Learning: ClueNet converts video into temporally grounded entity-interaction clues, learns which clues are both relevant to the question and supported by the frames, and uses two-stage supervision to reach 77.6%, 84.3%, and 69.8% on STAR, NExT-QA, and MVBench; optional visual compression accelerates inference but is a different operating point from the highest-accuracy configuration.
Constructing and Interpreting Digital Twin Representations for Visual Reasoning via Large Language Models and Reinforcement Learning: DT-R1 introduces a unified pure reinforcement learning (GRPO) framework that trains a single LLM to autonomously orchestrate vision foundation models into digital twin (DT) representations, reasoning over continuous geometric and temporal properties via explicit chains of thought and executable Python code to resolve diverse reasoning visual tasks without task-specific architectures or supervised fine-tuning.
Context-Interactive Reasoning for Group Activity Detection: To tackle occlusion-induced spurious associations and context noise in crowded multi-view videos, this paper proposes a Context-Interactive Reasoning framework that softly constrains spatial topology via STIR and adaptively conditions on global scene semantics via gated LRCC, establishing new state-of-the-art performance on GAD benchmarks.
CORE-V: Chain-Of-thought REasoning for Image Editing with Visual Interaction: CORE-V addresses the lack of visual awareness and spatial geometric grounding in existing text-based CoT image editing frameworks by introducing an explicit Visual Chain-of-Thought (Visual CoT) pipeline that dynamically synchronizes attribute visual imagination with adaptive multi-modal spatial conditions for fine-grained image synthesis.
Demystifing Video Reasoning: Refuting the prevailing assumption that video diffusion models reason sequentially across temporal frames (Chain-of-Frames), this paper discovers that reasoning fundamentally unfolds along diffusion denoising steps (Chain-of-Steps) via structured multi-path exploration and working memory, and presents a training-free latent ensemble (TFE) that consistently boosts reasoning performance.
Diffusion-Based Immersive Visual Reasoning: To overcome partial observability and rigid open-loop perception in 360° scenes, DIVR reframes immersive visual reasoning as a denoising diffusion process where a Latent Hypothesis Refiner and a Reasoning-Guided Observer iteratively couple belief updating with active viewport selection, significantly outperforming existing VLMs.
DoCoG: Mask-based Multi-Type Grounded Chain-of-Thought for Document QA: DoCoG presents a mask-based multi-type grounded step-wise Chain-of-Thought architecture for Document QA that couples a vision-language model with a promptable segmenter via a Grounding Interaction Module (GIM) and anchor-regularized DPO, delivering fine-grained pixel-level evidence alignment across textual and graphical document structures.
EgoMAN: Interaction-Structured Reasoning for Egocentric 3D Hand Trajectory Prediction: EgoMAN introduces an interaction-structured reasoning framework that decomposes egocentric 3D hand motion into stage-aware phases, bridging high-level multimodal reasoning and low-level Flow Matching motion experts via a compact 4-token Trajectory-Token Interface.
EGVLR: Evidence-Grounded Vision–Language Reinforcement for Anomaly Reasoning: Addressing language-prior shortcuts, contradictory reasoning traces, and hallucinated defect boxes on normal samples in industrial MLLMs, this paper proposes EGVLR, a reinforcement framework unified under an Evidence-Driven Diagnostic Protocol (EDDP) across progressive visual pre-alignment, knowledge-grounded instruction tuning, decoupled preference optimization, and box-guided segmentation rendering, achieving state-of-the-art anomaly reasoning while slashing spatial false positives.
Enhancing Embodied Reasoning and Grounding by Novel View Synthesis: SplatReasoner turns a Gaussian scene into visual memory that can be reframed, combining visibility-based retrieval, local novel view synthesis, and text-based judging with verification to give a VLM more useful evidence, raising GPT-4o LLM-Match on OpenEQA from 3D-Mem's 54.6 to 57.8.
Evaluating Reasoning Coherence in Video Generative Models with Text and Visual Hints: Proposes MME-CoF-Pro, the first video reasoning benchmark that explicitly decouples hint modalities and evaluates process-level intermediate steps, systematically revealing severe reasoning coherence deficits and prompt-induced hallucinations across 7 frontier video models.
Exploring Efficient Reasoning Segmentation with Small Language Models: Rather than relying entirely on a small model's single <SEG> token, LReSeg supplies dense spatial information through a unified visual encoder and text-guided register tokens, achieving 50.8 gIoU and 50.0 cIoU on the ReasonSeg test set with 805M parameters while reducing LISA-7B's latency from 320 ms to 98 ms under the paper's timing setup.
GR-GRPO: Graph-Diffused Credit for Autoregressive Image RL Alignment: Addressing the compute-heavy rollout bottleneck and point-wise credit sparsity in small-group (\(G\)) GRPO for VQ-based autoregressive image generation, GR-GRPO leverages local substitutability in codebook embedding space to diffuse positive rollout evidence across a precomputed K-NN graph into soft targets, outperforming large-group GRPO (\(G=64\)) with \(\sim 4.6\times\) less rollout time on Janus-Pro-1B.
GradingBench: Evaluating End-to-End Compositional Reasoning of MLLMs for Automated Exam Grading: Presents GradingBench, the first comprehensive benchmark evaluating end-to-end automated exam grading across 300 full-page K–12 exam sheets and 3,284 annotated sub-questions, demonstrating that visual localization is the primary bottleneck and that MLLMs suffer from severe coordination failures when integrating perception and reasoning.
Hi-Nav: Hierarchical Framework for Continuous Vision-Language Navigation via Map Guidance and Waypoint Reasoning: To bridge the persistent sim-to-real gap and mitigate compounding errors caused by coupling semantic reasoning with physical control in continuous vision-and-language navigation (VLN-CE), Hi-Nav proposes a top-down three-level hierarchical framework—integrating LLM-based long-horizon instruction decomposition, Q-Former-driven environment perception with distance-regularized waypoint reasoning, and ROS-compliant velocity execution—achieving an 80% Success Rate in Gazebo and robust real-world robot deployment.
HighlightBench: Benchmarking and Diagnosing Markup-Driven Table Reasoning in Scientific Documents: Addressing the evaluation blind spot where multimodal large language models cannot distinguish between perceiving visual cues and executing logical constraints, HighlightBench introduces a diagnostic benchmark of 446 tables and 3,283 questions across five decoupled task families, counterfactual perturbations, and an explicit DSL reference pipeline to expose how visual salience systematically misleads reasoning.
Human-Level Accuracy, Non-Human Strategies: Revealing Model-Human Divergence in Video Physical Reasoning: Addressing the question of whether video foundation models matching human-level accuracy genuinely simulate intuitive physics, this paper proposes a population-level distributional evaluation framework comparing model probe seeds against human rater distributions, revealing that models form a coherent yet distinct non-human strategy driven by visible statistical cues rather than explicit forward simulation.
i-Design: Step-by-Step Graphic Layout Design with Progressive Aesthetic Policy Optimization: Addressing the visual imbalance and spatial misalignment inherent to conventional single-shot layout models, i-Design reformulates 2D graphic layout generation as an iterative group-wise trajectory decision process conditioned on partially rendered intermediate canvases, integrating Progressive Imitation Learning (PIL) with online Progressive Aesthetic Policy Optimization (PAPO) driven by ground-truth anchored consensus graph rewards to achieve state-of-the-art geometric alignment and human aesthetic preferences.
Learning Active Perception for Pixel-Space Reasoning via Visual-Intent Stratified GRPO: Addressing the visual laziness failure mode and degradation of multi-zoom behavior when training pixel-space reasoning models with standard GRPO, this paper introduces Visual-Intent Stratified GRPO (VIS-GRPO), which stratifies rollout groups jointly by zoom depth and spatial-coherence-based visual intent (explore vs. verify) to ensure fair group-relative advantage normalization, boosting performance by 8.7% on HR-Bench 8K and 7.6% on MME-RealWorld.
Learning Consistency in Reward Modeling for Multi-Modal Reasoning: Consistency Reward Models (CRM) reframe multimodal reasoning reward evaluation into chain-of-thought (CoT) guided semantic and mathematical equivalence checking between candidate outputs and ground-truth references in text space, deriving continuous soft rewards coupled with advantage-sign reject sampling to unlock 4× training data and achieve state-of-the-art RL performance.
Learning from Primitive: Probing Visual Reasoning of LVLMs via Counting: This paper establishes visual counting as a minimalist probe to uncover the internal reasoning mechanisms of large vision-language models (LVLMs), introduces Visual Activation Patching (VAP) and HeadLens to identify a multi-stage counting circuit, and shows that lightweight fine-tuning on simple synthetic primitives significantly boosts out-of-distribution counting and general visual reasoning.
Less is More: A Simple yet Effective Object-Centric Prompting Strategy for Vision-Language Reasoning in Autonomous Driving: By superimposing detected 2D bounding boxes and discrete object IDs directly onto multi-view images and converting coordinate regression into discrete ID prediction followed by offline dictionary lookup, this paper fundamentally eradicates spatial localization hallucinations in autonomous driving VLMs.
Metric-Bench: Exploring In-context Spatial Metric Reasoning in VLMs for Indoor Scenes: To address the lack of physical scale understanding and the vulnerability of Vision Language Models (VLMs) to catastrophic forgetting under rigid pixel-level supervision, this paper presents Metric-Bench, the first indoor benchmark for in-context absolute metric reasoning without camera intrinsics, and introduces MetricReasoner, an RFT framework combining structured chain-of-thought prompting with verifiable continuous-discrete numerical rewards that outperforms leading commercial and open-source models while preserving general multimodal intelligence.
Mitigating Pose–Scale Discrepancy Bias and Reforming Multi-Support Reasoning for Few-Shot Semantic Segmentation: Addressing catastrophic matching failure caused by pose-scale discrepancies and coarse prototype averaging in few-shot semantic segmentation, DMNet introduces an equivariant geometry-decoupled autoencoder to isolate semantic identity from geometric variation, paired with cyclic ensemble refinement and spatially adaptive reliability weighting to achieve new state-of-the-art results on PASCAL-5i and COCO-20i.
Mixture of Specialized Vision Experts: Unlocking Complementary Visual Insights for Faithful MLLM Reasoning: Addressing perceptual blind spots in single-encoder MLLMs and feature redundancy from naive expert concatenation, MoSVE introduces query-guided patch-level adaptive complementary clustering to compress visual tokens, paired with Complementary Rejection Fine-Tuning (CRFT) that routes anchor failure samples to the most visually disparate auxiliary expert, achieving SOTA hallucination mitigation on POPE (86.9%), CHAIR_S (27.1%), and MMVP (57.0%) using only 576 visual tokens.
MonoSR: Open-Vocabulary Spatial Reasoning on Monocular Images: MonoSR constructs the first large-scale open-world monocular spatial reasoning dataset (1M+ QA pairs, covering indoor, outdoor, and object-centric scenes with 98 semantic categories). It ensures that every QA pair is answerable from a single RGB image through a four-stage observability filtering mechanism, and systematically evaluates the impacts of seven open/closed-source VLMs as well as various auxiliary signals on spatial reasoning capabilities.
Natural Image Pretraining Improves Abstract Reasoning: This paper introduces Nat-ARC, demonstrating for the first time that visual self-supervised pretraining via Masked Autoencoders (MAE) on natural images (ImageNet) transfers effectively to vision-centric solvers for few-shot abstract grid reasoning on the ARC benchmark, achieving 63.4% pass@2 with a single model and 70.2% pass@2 via multi-regime prediction-level ensembling.
PASR: Pattern-Aware Scene-Conditioned Reasoning for Camouflaged Object Detection: PASR introduces an annotation-free, zero-training framework for camouflaged object detection that reframes the task from foreground discrimination into scene-conditioned pattern-deviation reasoning over a frozen feature space, significantly outperforming existing unsupervised/weakly supervised methods and closing the gap to fully supervised models.
PercepTax: Benchmarking Cross-Property Reasoning in Vision-Language Models: PercepTax introduces a benchmark for hierarchical perceptual taxonomy and cross-property reasoning across material, affordance, function, and physical attributes, revealing that frontier vision-language models suffer substantial performance drops when jointly integrating multiple physical constraints.
PGCR: Pose–Geometry Coupled Reasoning for Image-to-Point Cloud Registration: Addressing the vulnerability of conventional overlap-then-registration pipelines where inaccurate overlap prediction irreversibly discards valid regions, PGCR re-conceptualizes 6-DoF image-to-point cloud registration as coupled pose-geometry reasoning, achieving 100% registration accuracy and 0.15 m / 0.44° precision on KITTI with a fast 0.17 s runtime.
Reflection-Aware Reasoning for Non-Line-of-Sight Pedestrian Localization: Addressing the severe multipath distortion and clutter caused by ego-vehicle dynamics, this paper proposes an end-to-end reflection-aware multimodal learning and physics-guided ray tracing framework that fuses camera and 2D mmWave radar representations in BEV space to disentangle higher-order reflections and reconstruct reflective surfaces, reducing NLOS pedestrian localization error to 1.23 m while providing a 0.77 s advance warning on a moving testbed vehicle.
Reinforcing Vision-Language Models for Image Quality Assessment with Grounding Process Rewards: Addressing the issue of flawed logical attributions and poor out-of-distribution generalization caused by sparse outcome-only rewards in no-reference image quality assessment (NR-IQA), this paper proposes RGPR, a reinforcement learning framework that orchestrates a "ground-before-assess" reasoning chain via grounding process rewards and step-wise reward normalization.
ROVA: Are Video Reasoning Models Ready to Go Outside?: ROVA proposes a training framework for robust video reasoning that enhances VLM performance under real-world perturbations (weather, occlusion, illumination, camera shake) through structured spatio-temporal perturbation generation, self-reflective difficulty-aware online curriculum, and dual-branch alignment based on GRPO. Concurrently, it introduces PVRBench—the first embodied video reasoning benchmark with systematically injected real-world perturbations, covering over 9K+ videos and 51K+ QA pairs.
SEERBench: A Spatial Ego-Exo Reasoning Benchmark for MLLMs with a Simple Yet Effective Baseline: To tackle the severe failure of multimodal LLMs in cross-view geometric deduction between egocentric and exocentric perspectives, this paper introduces SEERBench, a benchmark spanning 3 cognitive levels and 10 tasks, reveals a stark 46.8% accuracy gap behind human experts, and proposes SEER-Map, a training-free bird's-eye view topological prior that substantially eases multi-view spatial reasoning.
Spatial Amsan: A Benchmark for Perception-Grounded Spatial Reasoning and Action Evaluation in Egocentric Manipulation: Introduces Spatial Amsan, the first embodied spatial mental arithmetic benchmark for egocentric manipulation videos across two tasks (Tangram and Wooden Puzzle, totaling 282 scenarios and 1,801 evaluation turns); through a diagnostic Q1–Q4 protocol, it reveals that frontier VLMs exhibit a severe dissociation between stable perception and sharply degrading action evaluation over time.
STEP: Spatial Thinking and Egocentric Pointing for Embodied Instruction Following: STEP integrates global Bird's-Eye-View (BEV) semantic maps with egocentric visual waypoint prompting, driven by an automated 400k+ multimodal Chain-of-Thought data engine to establish an interpretable point-level planning paradigm: "Look at the View, Think with the Map, and Point to the Goal."
StreamSpatial: A Benchmark and Framework for Streaming 3D Visual-Spatial Reasoning: Addressing the unrealistic assumption of static global spatial omniscience in prior benchmarks, this paper introduces StreamSpatialBench with 5,280 millisecond-grounded QA pairs across 2,003 videos, alongside Stream3D-Mem, a lightweight streaming framework unifying bounded spatiotemporal memory consolidation with latent cognitive map reasoning.
TerrainGraphNet: Terrain-Constrained Graph Reasoning for Landslide Segmentation: Addressing the limitation of conventional landslide segmentation where Digital Elevation Models (DEM) are treated as generic appearance channels without geomorphological constraints, TerrainGraphNet introduces terrain-modulated feature interaction and slope-constrained sparse graph reasoning to suppress boundary leakage and topological fragmentation.
Thinking from the Robot’s View: The CoT-HRC Benchmark for Human Intent Reasoning in Embodied Collaboration: To tackle spatiotemporal incompleteness and black-box evaluation pitfalls in human intent inference from realistic quadruped robot views, this paper introduces CoT-HRC in Habitat 3.0, pairing low-angle physical occlusion modeling with hierarchical Chain-of-Thought annotations to rigorously decouple perception failure from cognitive reasoning.
Toward Interpretable Analysis of Whole-slide Pathology Images via Large Language Model-based Agentic Reasoning: PathAgent is a training-free LLM-based agent framework for whole-slide pathology image (WSI) analysis that emulates the iterative, reflective inspection workflow of human pathologists using a Navigator, a Perceptor, and an Executor, establishing fully traceable and clinically grounded evidence trajectories.
UniReflect: Self-Reflection Tuning for Unified Multimodal Understanding and Generation: UniReflect presents the first unified multimodal framework that integrates visual comprehension, generation, and in-model verification, utilizing a shared Semantic Perceiver and a decoupled verification strategy to realize autonomous test-time scaling (TTS) without external discriminators.
Unlocking Complex Image Editing via Natively Interleaved Visual Textual CoT with Deep Confidence Reasoning: MURE proposes a unified multimodal framework that natively interleaves textual reasoning with intermediate visual rationales (precise masks, new object concepts) in an autoregressive stream, reinforced by Multimodal Deep Confidence (MMDC) tree pruning to achieve state-of-the-art complex image editing.
Vero: Open Reinforcement Learning Recipes for Visual Reasoning: Addressing the lack of open data and reward designs in leading vision-language models, Vero introduces the first fully open single-stage multi-task reinforcement learning recipe across six visual reasoning categories, building Vero-600K and task-routed rewards that outperform distillation-heavy open models across 30 benchmarks.
VisWordBench: Bridging the Gap in Cross-modal Reasoning for Multimodal Large Language Models: The paper introduces VisWordBench, a bilingual benchmark consisting of 4,625 densely annotated visual word puzzles, reveals that the core bottleneck of current MLLMs in cross-modal reasoning stems from under-diversified hypothesis search, and proposes a multi-stage synthetic data pipeline along with a training-free inference-time activation steering method.
VLTR: Vision-Language Tool Reasoning for Instruction-Guided Image Editing: VLTR is a training-free vision-language tool reasoning framework that reformulates instruction-guided image editing into closed-loop search over a directed acyclic graph (DAG) of atomic primitives. By marrying Bayes-UCB adaptive tool routing with BLUE-style heteroscedastic multi-tier verification and localized DAG repair, it achieves the best overall VLM ranking on PIE-Bench++ while offering a 4.8× runtime speedup over previous multi-tool systems on MagicBrush.
Wan-R1: Verifiable-Reinforcement Learning for Generalizable Video Reasoning: Wan-R1 adapts Flow-GRPO to flow-matching video generation and establishes verifiable trajectory-grounded and embedding-level reference-anchored rewards, overcoming multimodal reward hacking and substantially advancing out-of-distribution spatial reasoning in complex 3D mazes and robotic path planning.
Wavelet-based Intra-video Counterfactual Reasoning for Video Question Grounding: To tackle intra-video temporal bias caused by visually identical yet non-causal frames in weakly-supervised Video Question Grounding, WICI introduces wavelet-based dynamic modeling to capture fine-grained motion cues and curriculum-guided intra-video counterfactual mining with margin-based supervision for faithful evidence localization and robust reasoning.
YesTrack: Referring Multi-Object Tracking via MLLM-based Yes/No Reasoning: YesTrack does not ask the MLLM to generate target descriptions; it directly reads the logits of the Yes/No decision tokens in the vocabulary for discriminative referring verification, adds two inference-time constraints—Temporal Confidence Prior (TCP) and Temporal Reference Propagation (TRP)—and achieves SOTA with the smallest Qwen3-VL-2B-Instruct: HOTA 54.00 on Refer-KITTI and 43.75 on Refer-KITTI-V2, while the same paradigm transfers into a generic MOT tracker, YesTrack-MOT (KITTI HOTA 45.10).