🧠 VLM Reasoning¶
🧠 NeurIPS2026 · 81 paper notes
📌 Same area in other venues: 🎞️ ECCV2026 (164) · 📷 CVPR2026 (150) · 🔬 ICLR2026 (112) · 💬 ACL2026 (32) · 🧪 ICML2026 (31) · 🤖 AAAI2026 (10)
🔥 Top topics: Reasoning ×74 · Multimodal/VLM ×31 · Face & Gaze ×3 · LLM ×3 · Agents ×3
- 2D Spatial Reasoning with Adaptive Neural Cellular Automata
-
Addressing the failure of classical Neural Cellular Automata (NCAs) on long-range geometric dependencies caused by rigid local receptive fields, this paper introduces Adaptive Neural Cellular Automata (aNCA), which leverages multi-head modulated deformable convolutions to dynamically discover and reason over spatial constraints with tiny parameter counts, setting new state-of-the-art results on Sudoku and maze navigation.
- A Benchmark for Omni-Modal Reasoning in Long Videos
-
LongShOTBench introduces an omni-modal long-video benchmark covering 274 hour-scale videos (avg. 41 mins), 3,401 intent-driven QA samples across single/multi-turn dialogues with weighted criterion-level rubrics, alongside LongShOTAgent, a training-free search-refine-verify agent that achieves the state-of-the-art score of 66.64% across 105 evaluated systems.
- A Systematic Evaluation of Vision-Language Models for Observational Astronomical Reasoning Tasks
-
Introduced AstroVLBench, a multimodal astronomical reasoning benchmark comprising over 4,100 expert-verified instances across five observational modalities, systematically revealing that frontier VLMs exhibit acute modality dependence, severe bottlenecks in quantitative grounding, and a prevalent "right-answer-wrong-reason" decoupling.
- AgroOmni: A Large-Scale Multi-View Agricultural Dataset for Cross-Scale Multimodal Reasoning
-
To eliminate the severe ground-level bias and scale confusion in agricultural multimodal models, AgroOmni establishes a 288K-pair multi-view VQA dataset bridging Ground, UAV, and Satellite perspectives, enabling the proposed AgroNVILA baseline to achieve a 62.32% SOTA on AgroMind (+15.03% over GPT-5.2).
- Allocentric Perceiver: Disentangling Allocentric Reasoning from Egocentric Visual Priors via Frame Instantiation
-
Allocentric Perceiver is a training-free, plug-and-play framework that leverages off-the-shelf 2D/3D geometric models to recover metric states, dynamically instantiates query-conditioned allocentric reference frames, and prompts VLMs with structured geometric text, successfully offloading mental rotation into explicit computation.
- AlloSpatial: Agentic Harness Framework for Spatial Reasoning in Foundation Models
-
AlloSpatial introduces an agentic harness framework that translates egocentric video observations into queryable Allocentric-Spatial Trees (ASTs) via the World2Mind sandbox, internalizing decoupled evidence collection and cross-modal arbitration through harness-gated reinforcement learning.
- ANCHOR: Audio-Visually Grounded Chain-of-Thought Reasoning Benchmark
-
Introduced ANCHOR, an audio-visual benchmark of 2,319 videos and 8,272 QA pairs integrating SAM3 spatial bounding boxes, temporal intervals, acoustic cues, and human-refined rationales, revealing that models achieving comparable answer correctness frequently exhibit severe perceptual unfaithfulness and reasoning divergence.
- ARCANA: A Benchmark for Abstraction and Analogical Reasoning
-
To evaluate far analogical reasoning between superficially disparate domains, the paper introduces ARCANA, a 200-task benchmark mapping physical principles and real-world objects onto abstract discrete grids, uncovering a significant human-LLM cognitive gap, and proposes the ConceptSIR framework to overcome reasoning local minima via concept slippage, code execution, introspection, and reflection.
- ARK: A Dual-Axis Multimodal Retrieval Benchmark along Reasoning and Knowledge
-
Presents ARK, the first multimodal retrieval benchmark that explicitly decouples knowledge domains from reasoning skills across 17 subtypes and 6 reasoning dimensions, revealing severe bottlenecks of current retrievers in fine-grained visual and spatial reasoning.
- ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both
-
ATLAS internalizes complex visual operations—such as geometric auxiliary lines, region zooming, and symbolic tagging—into single discrete vocabulary tokens without pixel generation or external tool execution, addressing sparse token gradient dilution via Latent-Anchored GRPO (LA-GRPO).
- AVIS: Adaptive Test-Time Scaling for Vision–Language Models
-
To tackle the severe memory and compute inflation caused by visual context prefilling and long chain-of-thought generation in vision-language models, AVIS introduces a training-free \(O(N)\) Key Diversity Visual (KDV) pruning rule coupled with a lightweight difficulty predictor under shared-prefill inference, substantially cutting FLOPs and latency while outperforming the vanilla baseline across diverse multimodal benchmarks.
Benchmarking Multimodal Mathematical Reasoning with Explicit Visual Dependency
- BrainTRACE: Tracing Longitudinal, Multimodal, and Volumetric Evidence in Brain MRI Clinical Reasoning
-
BrainTRACE is the first report-grounded evaluation benchmark centered on longitudinal brain MRI follow-up and clinical reasoning, featuring 7,273 structured VQA instances across 1,778 longitudinal patients and 7,299 MRI studies, exposing a severe evidence attrition gap where frontier VLMs excel at isolated recognition but collapse during full-case longitudinal synthesis.
- Bridging Modalities, Spanning Time: Structured Memory for Ultra-Long Agentic Video Reasoning
-
For ultra-long video understanding spanning days to weeks, MAGIC-Video introduces a training-free framework that combines a Multimodal Memory Graph and a Narrative Memory Chain, enabling cross-modal retrieval via Personalized PageRank and top-down cross-time fact injection for multi-round agentic reasoning.
- Can Linguistic Reasoning Vectors Enhance Multimodal Reasoning Ability?
-
LIFT extracts the hidden-state difference at identical answer tokens with and without a reasoning trace from a base large language model, then injects it into the corresponding vision-language model's language layers without updating the backbone, raising six-benchmark average accuracy from 66.6/67.8 to 68.3/69.0 with static intervention and 69.1/69.4 with vector adaptation.
- CausalSpatial: A Benchmark for Object-Centric Causal Spatial Reasoning
-
This paper introduces CausalSpatial, the first object-centric benchmark diagnosing causal spatial reasoning in MLLMs, revealing severe hallucinations in textual chain-of-thought, and proposes COW, a 4D trajectory-controlled world model that externalizes physical dynamics simulation into video frames.
- CODA: Cohort- and Drift-Aware Foundation Model for Multimodal Clinical Reasoning
-
Addressing the limitation of conventional medical vision-language models that reason solely at the isolated instance level, CODA establishes a two-stage paradigm combining reinforcement learning-based hierarchical cohort discovery with prototype drift-aware individual adaptation, setting new state-of-the-art performance across multimodal clinical QA and report generation benchmarks.
- CoLVR: Enhancing Exploratory Latent Visual Reasoning via Contrastive Optimization
-
To overcome latent space homogenization and exploratory collapse caused by rigid feature alignment in latent visual reasoning, CoLVR introduces angle-perturbed latent contrastive fine-tuning alongside a trajectory-level contrastive reward in reinforcement learning, substantially expanding representation diversity and multi-step reasoning capabilities.
- Distinguishing Performance From Competence in Evaluations of Humanlike Abstract Reasoning
-
By evaluating frontier reasoning models on the ConceptARC benchmark across text and vision modalities and auditing their generated natural-language rules, this paper reveals that benchmark accuracy significantly overestimates AI competence in text by rewarding spurious shortcuts, while substantially underestimating underlying competence in vision due to perceptual hurdles.
- DocScope: Benchmarking Verifiable Reasoning for Trustworthy Long-Document Understanding
-
DocScope formulates long-document QA as a four-stage structured reasoning trajectory prediction task (pages, bounding boxes, facts, answer) and shows that even among correct answers, the complete evidence-chain rate tops out at only 29%, demonstrating that end-to-end accuracy cannot proxy for reasoning verifiability.
- Does Seeing More Mean Knowing More? Mono-Anchored Advantage Normalization for Multi-Source Visual Reasoning
-
Addressing the issue where blind multimodal fusion in visual reasoning leads to modality interference and underperformance relative to mono-source inputs, this paper introduces MARS, a framework that leverages mono-source rewards as dynamic anchors in advantage normalization to adaptively reward information gain and suppress modality conflicts.
- DRAGON: A Benchmark for Evidence-Grounded Visual Reasoning over Diagrams
-
Introduces DRAGON, a multi-domain diagram evidence grounding benchmark spanning 2,445 human-verified instances across 6 diverse domains, proposing EDGE, SAGE, and VERGE prompting protocols to reveal that leading VLMs exhibit severe visual evidence localization deficits even when predicting accurate answers.
- DRScaffold: Boosting Dense-Scene Reasoning in Lightweight Vision Language Models
-
To fix the systematic breakdown of lightweight vision-language models on dense-scene reasoning with multiple objects and relations, this paper introduces the DRBench benchmark and the DRScaffold supervised fine-tuning framework, which enforces grounded reasoning via a four-stage causally ordered schema and staged gradient masking, allowing a 3B model to outperform an un-tuned 32B model on dense-scene tasks.
- ElasticFit: Fit-Aware 3D Object Insertion via VLM Reasoning and Generative Adaptation
-
ElasticFit introduces a VLM-guided framework for fit-aware 3D object insertion that maps natural-language instructions and rendered scene views into structured geometric fitting cues, generating scene-conditioned object priors and refining them through rigid, uniform, or elastic adaptation alongside physics-based settling to achieve physically grounded placements in constrained spaces.
- Enhancing Gaze Reasoning in Vision Foundation Models for Gaze Following
-
Addressing the vulnerability of vision foundation models that exploit "semantic shortcuts" toward salient objects in gaze following, this paper introduces Head-Conditioned Local LoRA (HCLoRA) and an Out-of-Cone (OOC) geometric penalty to selectively enhance head gaze reasoning while preserving frozen scene semantics, achieving state-of-the-art accuracy particularly in non-salient target scenarios.
- EventPrune: Cascaded Event-Assisted Token Pruning for Efficient First-Person Dynamic Spatial Reasoning
-
Addressing the quadratic attention bottleneck of high-resolution Video-LLMs in first-person dynamic spatial reasoning, this paper proposes Event Cascade Pruning (ECP)—a training-free framework combining temporal causal sampling, motion saliency filtering, and rank-fused attention calibration to cut 80% visual tokens while surpassing full-token accuracy with 1.89× speedup.
- Evidence-RL: Towards Evidence-Intensive Visual Reasoning
-
Proposes Counterfactual Evidence Disentanglement (CED), a post-training audit that measures candidate answer support drop under feature-space mean replacement of object-centric evidence regions versus matched non-evidence regions, integrating the resulting causal evidence margin into GRPO to eliminate language-prior shortcuts with zero inference overhead.
- Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models
-
Addressing the severe norm mismatch and subspace shift between decoder hidden states and input vision embeddings in pre-norm MLLMs, GAP introduces a granular alignment paradigm spanning a PCA-aligned latent head, auxiliary visual context supervision, and difficulty-aware assignment, delivering robust perception and multi-hop visual reasoning gains without external tools.
- FinMTM: A Multi-Turn Multimodal Benchmark for Financial Reasoning and Agent Evaluation
-
Addressing the critical omission of multi-turn interaction, long-context reasoning, and autonomous agent workflows in existing single-turn financial benchmarks, FinMTM introduces a bilingual benchmark of 11,133 QA pairs across objective questions, cognitive-graded multi-turn dialogues, and MCP-based financial agent tasks, revealing persistent bottlenecks across 22 leading VLMs.
- From Document Layout to Causal Topology: An End-to-End Architecture for Temporal Causal Reasoning
-
Addressing the limitation of large language models in remaining confined to shallow timeline ordering rather than deep causal deduction, this work draws inspiration from visual document layout reordering to propose a training-free neuro-symbolic framework, TCR-Agent, and an end-to-end model, TCR-Former, that natively internalizes causal-topological attention masks and temporal span biases into the Transformer forward computation.
- GeoSym127K: Scalable Symbolically-Verifiable Synthesis for Multimodal Geometric Reasoning
-
This paper introduces GeoSym, a neuro-symbolic synthesis engine that unifies arbitrary-precision symbolic manifolds, CCA-based visual grounding, and analytic solvers to generate GeoSym127K and GeoSym-Bench; combined with SFT and deterministic exact-match GRPO, it significantly mitigates visual hallucination and long-horizon logic fragmentation in multimodal geometric reasoning.
- Gold Panning: Strategic Context Shuffling for Needle-in-Haystack Reasoning
-
Addressing severe "lost-in-the-middle" position bias in long-context LLMs, Gold Panning introduces a black-box Bayesian active search framework that pairs position diagnosticity calibration with signal anchoring to match Permutation Self-Consistency retrieval accuracy while cutting query calls by 30%–65%.
- Hear, Localize and Reason: Spatially Aware Scene Understanding for Audio-Visual LLMs
-
Addressing the limitation of conventional audio-visual LLMs that treat audio as monaural semantic content without spatial perception, this paper establishes a hear-localize-reason benchmark (HLR-AVSceneQA) and proposes HLR-LLM, augmenting a foundation model with a binaural spatial audio branch and staged chain-of-thought fine-tuning to enable robust spatial grounding and relational scene understanding.
- Hide to See: Reasoning-Prefix Masking for Visual-Anchored Thinking in VLM Distillation
-
Addressing the visual forgetting issue where compact VLMs over-rely on accumulated textual prefixes during long think-answer distillation, Masking-KD selectively masks salient reasoning prefixes and adaptively schedules the masking budget, forcing the student to exploit visual evidence and establishing new state-of-the-art multimodal reasoning across 2B, 4B, and 8B scales.
- HopChain: Multi-Hop Data Synthesis for Generalizable Vision-Language Reasoning
-
HopChain proposes a scalable multi-hop data synthesis framework for reinforcement learning with verifiable rewards (RLVR), enforcing continuous visual re-grounding across long reasoning chains and boosting performance on 20 out of 24 multimodal benchmarks.
- Imagine3D-LLM: Teaching MLLMs to Imagine 3D Scenes before Answering
-
Imagine3D-LLM equips multi-image MLLMs with the ability to perform mental 3D reconstruction before answering by decoding compact 3D Gaussian summary tokens under an information bottleneck, achieving new state-of-the-art accuracy across seven 3D spatial reasoning benchmarks.
- Learning to Evolve Scenes: Reasoning about Human Activities with Scene Graphs
-
To overcome the limitations of implicit continuous representations in egocentric video understanding, this paper introduces SG-Ego, a large-scale spatio-temporal scene graph dataset, and GLEN, a graph edit network that casts human-action-driven scene evolution as an activity-conditioned graph edit forecasting (A-GEF) task, enabling explicit, structured, and interpretable dynamic reasoning.
- Look Before You Reason: Implicit Visual Thinking for Efficient Multimodal Reasoning
-
This paper proposes Look Before You Reason (LoRe), an implicit visual thinking framework that autonomously predicts task-relevant regions and amplifies their corresponding KV cache Value entries within a single VLM generation turn via GRPO reinforcement learning, boosting fine-grained visual reasoning with negligible latency overhead.
- MMOU: A Massive Multi-Task Omni Understanding and Reasoning Benchmark for Long and Complex Real-World Videos
-
To address the lack of long-horizon joint audio-visual reasoning evaluation in existing MLLMs, this paper proposes MMOU, a massive omni-modal understanding benchmark comprising 20,000 expert-annotated multiple-choice questions and open-ended QAs across 11,877 real-world long videos (up to 121 minutes) and 13 fundamental reasoning skills, exposing a severe capability gap where the strongest closed-source model achieves only 64.2% accuracy and the best open-source model reaches just 46.8%.
- More Than Meets the Eye? Uncovering the Reasoning-Planning Disconnect in Training Vision-Language Driving Models
-
This paper exposes a pervasive "reasoning-planning disconnect" in vision-language driving agents, demonstrating that trajectory planning relies heavily on textual prior shortcuts rather than the generated chain-of-thought reasoning, and introduces a training-free causal probe to diagnose this shortcut reliance.
- Motion-O: Trajectory-Grounded Video Reasoning
-
Addressing the failure of video vision-language models to ground inter-frame dynamics, Motion-o introduces Spatial-Temporal-Trajectory (STT) reasoning with discrete Motion Chain-of-Thought (MCoT) tags, optimized via trajectory consistency and temporal perturbation rewards without architectural modifications.
- MVVBench: Benchmarking 4D Reasoning in Vision-Language Models
-
Addressing the absence of joint spatial-temporal multi-view evaluation in multimodal systems, this paper introduces MVVBench, the first 4D multi-view video reasoning benchmark comprising 1,323 expert-crafted QA pairs strictly governed by a monocular-ambiguity constraint, and presents training-free elicitation frameworks including Multi-View-Guided CoT and Tool-Augmented Dynamic Scene Graphs (TADSG) that expose and mitigate severe cross-view correspondence bottlenecks in current VLMs.
- NeuronEye: Query-Guided Visual Concept Activation for Vision-Language Reasoning
-
NeuronEye constructs a sparse, concept-level neuron vocabulary from intermediate VLM representations using a sparse autoencoder, dynamically activates query-relevant visual evidence via language-guided routing, and suppresses dominant perceptual directions at deeper layers, delivering substantial gains on multi-view and spatial relational reasoning without fine-tuning the backbone.
- NoRA: Evaluating Grounded Reasonableness in Visual First-person Normative Action Reasoning
-
Introduces NoRA, a visual first-person video benchmark that moves beyond multiple-choice shortcuts by evaluating normative action reasoning via explicit fact-reason-action support graphs and a geometric grounded reasonableness metric (\(R\)), exposing that current frontier VLMs struggle to construct the full reasonable action space and bind actions to appropriate local support.
- OmniJigsaw: Enhancing Omni-Modal Reasoning via Modality-Orchestrated Reordering
-
Addressing the lack of scalable self-supervised reinforcement learning proxy tasks for audiovisual omni-modal models, OmniJigsaw formulates temporal jigsaw reordering over video and audio chunks, identifies the bi-modal shortcut phenomenon, and introduces Chunk-level Modality Masking (CMM) alongside a two-stage coarse-to-fine filtering pipeline to enforce cross-modal information bottlenecks and substantially boost multimodal reasoning.
- On the Faithfulness of Visual Thinking: Measurement and Enhancement
-
Addressing the widespread unfaithfulness in multimodal chain-of-thought (MCoT) where models generate flawed visual evidence yet yield correct answers, this paper establishes a causal intervention and automated reliability/sufficiency evaluation pipeline, and introduces Sufficient-Component Cause Model (SCCM) learning to enforce minimal yet sufficient visual components via multiplicative gated RL rewards.
- OneCanvas: 3D Scene Understanding via Panoramic Reprojection
-
OneCanvas lifts multi-view image patch features into 3D and places them as separate continuous tokens in a shared panoramic coordinate system, using native RoPE for angles and frame order and an additive embedding for metric position, then combines synthetic spatial pretraining with real-scene QA adaptation to score 65.3, 71.3, and 72.1 on SQA3D, VSI-Bench, and SPBench, respectively.
- Perceive, Interact, Reason: Building Tool-Augmented Visual Agents for Spatial Reasoning
-
To overcome the failure of static vision encoders to recover fine-grained spatial evidence, this paper proposes PERIA, a tool-augmented visual agent that decouples visual perception from visual interaction tools, and develops OR-GIGPO for credit assignment over multi-step visual trajectories.
- PRISM: A Benchmark for Programmatic Spatial-Temporal Reasoning
-
Addressing the critical blind spot where language models synthesize runnable animation code that suffers from severe layout breakdowns, PRISM introduces a human-calibrated benchmark of 10,372 bilingual instruction-code pairs and a funnel-style deterministic geometric evaluation framework, uncovering an average 41% drop from execution success to spatial pass rate (the Execution-Spatial Gap).
- Reason in the Words You Speak: Idiolectal Paraphrasing Off-Policy Traces for Reasoning Distillation in VideoLLMs
-
Addressing the failure of mixed-policy GRPO where out-of-distribution teacher traces trigger trust-region clipping on semantically critical tokens, this paper proposes Echo-GRPO, which rewrites privileged traces into the student policy's native idiolect via Dual-Reference Decoding to enable stable and effective video reasoning distillation.
- RECAP: Looking Once Is Not Enough for Vision-Language Reasoning
-
Addressing the temporal grounding failure where long chain-of-thought VLMs drift from visual evidence under sparse terminal rewards, RECAP harnesses detached visual revisit traces to steer dynamic discounting and gated value prediction, substantially improving reasoning accuracy and mitigating hallucination without any inference-time overhead.
- Reinforcing Multimodal Reasoning Against Visual Degradation
-
Addressing the severe brittleness and reward poisoning that occur when reinforcement learning fine-tunes multimodal large language models under blur, noise, and compression artifacts, ROMA introduces a dual-forward-pass framework with teacher-forcing re-evaluation, correctness-conditioned worst-case KL regularization, and clean-advantage-anchored auxiliary policy gradients to boost visual degradation robustness without sampling on corrupted inputs.
- Seeing Together: Multi-Robot Cooperative Egocentric Spatial Reasoning with Multimodal Large Language Models
-
Addressing the bottleneck of cooperative spatial reasoning across multi-robot first-person video streams, this paper introduces CoopSR (a benchmark of 114K QA pairs) and SG-CoR, an architecture combining temporal frequency energy frame sampling, uncertainty-aware multi-view fusion, and privileged pose prompt distillation to substantially boost multi-robot reasoning without test-time pose access.
- Self-Rewarded Multimodal Coherent Reasoning Across Diverse Visual Domains
-
To resolve multimodal chain-of-thought hallucination and logical drift caused by outcome-only supervision, SR-MCR unifies five process-level intrinsic signals into an adaptive reliability-weighted self-reward, aligning vision-language reasoning via critic-free GRPO with confidence-aware cooling without human preference labels.
- SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning
-
Addressing the rigidity of single-pass code synthesis and structured JSON tool-calls in spatial agents, SpatialClaw establishes code execution within a persistent Python kernel as the agent action interface, empowering VLMs to iteratively compose perceptual and geometric primitives, inspect intermediate feedback, and dynamically refine spatial reasoning across 20 benchmarks.
- StemBind: When MLLMs Get Lost Between Rules and Instances in Abstract Visual Reasoning
-
Addressing the prevalent disconnect where multimodal large language models recognize elements and correctly state rules yet pick incorrect answers in abstract visual reasoning, this paper presents StemBind—the first shared-stem diagnostic benchmark combining Sternberg's componential cognitive framework with stage-wise stimulus interventions—identifying rule-to-instance mapping (S3 Map) as the primary bottleneck that neither scaling nor thinking mode can resolve.
- StreamGaze: Gaze-Guided Temporal Reasoning and Proactive Understanding in Streaming Videos
-
StreamGaze is the first benchmark designed to evaluate how multimodal large language models leverage human gaze dynamics for online streaming video understanding across past reflection, present perception, and proactive future alerts, revealing a major capability gap between state-of-the-art MLLMs and human performance (0.535 vs. 0.827).
- Structure Over Scale: Learning Visual Reasoning from Pedagogical Video
-
Addressing the vulnerability of vision-language models (VLMs) on fundamental spatial and navigation reasoning despite massive pre-training data, this paper presents the SoSVQA benchmark and an RL fine-tuning framework that extracts 10K question-answer pairs aligned with the context-question-pause-answer structure of preschool educational television, achieving substantial cross-benchmark gains (e.g., +19.70 on NExT-QA) using only 78 hours of video with GRPO.
- SVoT: State-aware Visualization-of-Thought for Spatial Reasoning via Reinforcement Learning
-
Addressing the lack of verification and implicit state transitions in MLLM multi-hop spatial planning, this paper introduces State-aware Visualization-of-Thought (SVoT), an RL framework that couples structured intermediate states with transition reasoning chains and visual synthesis, delivering up to a 65% absolute accuracy boost under GRPO-guided process-level optimization.
- Taming CoT Obfuscation in VLMs: From Mechanistic Evidence to Activation-Level Enforcement
-
Addressing the CoT obfuscation failure where reinforcement learning causes vision-language models to maintain high task accuracy while producing ungrounded, unmonitorable reasoning traces, this paper uncovers internal representational collapse between template and visual features and proposes TAME, an intervention framework combining anti-obfuscation feedback distillation with asymmetric SAE activation constraints.
- TaskGround: Structured Executable Task Inference for Full-Scene Household Reasoning
-
To overcome the fragility of household agents when facing cluttered complete scenes and situated, informal user requests, this paper introduces TaskGround—a training-free Ground–Infer–Execute framework that slashes input token costs by up to 18× while enabling compact open-weight models to rival frontier proprietary models on the newly established 400-task FullHome benchmark.
- Temporal Backtracking Search for Test-time Generative Video Reasoning
-
Addressing the fundamental bottlenecks that searching over denoising steps fails to rescue logically flawed rollouts and root-level Best-of-N blindly discards upstream verified progress, this paper introduces Temporal Backtracking Search (TBS). By integrating variable-\(K\) conditioning, temporal process verification, and prefix-priority search, TBS reformulates video reasoning as an iterative generate–verify–restart loop, dramatically outperforming matched-budget Best-of-N on long-horizon and out-of-distribution reasoning benchmarks.
- Test-time scaling of diffusions with flow maps
-
This paper introduces Flow Map Trajectory Tilting (FMTT), an inference-time adaptation framework that leverages flow maps for accurate look-ahead terminal reward evaluation and gradients, collapsing Jarzynski importance weights into an elegant formula and significantly outperforming standard diffusion test-time guidance across geometric and vision-language model rewards.
- Think, Then Score: Decoupled Reasoning and Scoring for Video Reward Modeling
-
Addressing the credit assignment bottleneck and high-variance gradient instability in coupled generative video reward models, DeScore decouples multimodal CoT reasoning from continuous scalar scoring via a dedicated query token and dual-objective RL, outperforming prior SOTA with 76% less training data.
- Thinking with Imagination: Agentic Visual-Spatial Reasoning with World Simulators
-
This paper presents Astra, an agentic spatial reasoning framework that couples an action-conditioned world simulator (Astra-WM) with a two-phase RL-trained VLM policy (Astra-VL), enabling models to actively query and ground imagined novel-view observations to resolve spatial ambiguity.
- ThinkJEPA: Empowering Latent World Models with Large Vision–Language Reasoning Models
-
ThinkJEPA couples a dense-frame V-JEPA latent dynamics branch with a uniformly sampled Qwen3-VL thinker branch via a dual-temporal pathway, distilling multi-layer reasoning representations through a hierarchical pyramid module and layer-wise FiLM modulation to break the bottleneck of semantic-free latent forecasting across egocentric manipulation, robotic rollout, and physical scene benchmarks.
- TopoGraphRAG-Bench: Evaluating Multimodal GraphRAG on Layout-Grounded Evidence Reasoning
-
Addressing the challenge that real-world document evidence is dispersed across text, tables, figures, and complex layouts, TopoGraphRAG-Bench is introduced as the first benchmark explicitly evaluating multimodal evidence topology recovery, using 2,024 counterfactually validated QA instances across single-hop, bridge-chain, and synthesis topologies to systematically expose failure modes of text-only GraphRAG, visual page retrieval, and multimodal GraphRAG.
- Touch-R1: Reinforcing Touch Reasoning in MLLMs
-
Touch-R1 establishes a large-scale tactile reasoning benchmark TouchReason-1M across four optical tactile sensors and introduces a tactile-grounded GRPO reinforcement learning objective combining ordinal accuracy, cross-sensor consistency, structured formatting, and counterfactual tactile grounding to elicit verifiable physical reasoning in MLLMs.
- Train the Agent, Not the Expert: Learning to Harness Heterogeneous Experts for Multi-Turn Visual Reasoning
-
VisHarness decouples high-level perception and decision-making from low-level execution, introducing dynamic visual memory archiving and online GRPO reinforcement learning so that a lightweight-trained visual agent learns to autonomously harness heterogeneous visual experts across complex segmentation, detection, and counting tasks.
- UniPath: Adaptive Coordination of Understanding and Generation for Unified Multimodal Reasoning
-
UniPath introduces a planner-executor framework that formalizes multimodal reasoning into the adaptive selection and execution of 5 distinct coordination paths, coupling aligned visual-thought supervision for the executor with query-form calibrated planning to achieve a Pareto improvement in reasoning accuracy and token efficiency.
- UniRAP: Towards Unified Part-level Physical Affordance Reasoning and Actionable Perception
-
UniRAP unifies multimodal prompt modeling and actionable perception by coupling Qwen2.5-VL with SAM 2 via interface tokens and a Unified Affordance Decoder, predicting part-level affordance masks, bounding boxes, and 4-DoF grasp poses across static images, videos, and real-world robot execution.
- UniVR: Thinking in Visual Space for Unified Visual Reasoning
-
Addressing the fundamental limitation of traditional MLLMs that rely on text-space chains-of-thought and struggle to capture continuous physical dynamics, UniVR presents the first unified framework that bypasses language intermediaries to reason natively in visual space, powered by VR-GRPO reinforcement learning with complementary global and step-focal rewards, achieving up to 25% improvements on the comprehensive VR-X benchmark.
- Unveiling Fine-Grained Visual Traces: Evaluating Multimodal Interleaved Reasoning Chains in Multimodal STEM Tasks
-
Introduces StepSTEM, a graduate-level benchmark of 283 challenging problems with strict text-vision complementarity, alongside a fine-grained evaluation framework coupling monotonic dynamic programming and patch-based visual alignment to rigorously assess multimodal interleaved reasoning chains.
- VICO: Visual Environments Co-Evolving for Vision-Language Model Reasoning
-
Addressing the learning signal collapse in post-training VLMs caused by static visual environments drifting out of the actor's learning frontier, VICO introduces an actor-environment co-evolution framework where an EnvRewriter edits structured visual representations (scene graphs, data tables, region masks) and deterministically re-renders them, dynamically steering task difficulty toward the actor's learning frontier (\(\hat{p} \approx 0.5\)) via rollout pass-rate rewards without extra human annotation.
- VISD: Enhancing Video Reasoning via Structured Self-Distillation
-
To tackle sparse sequence rewards and lack of token-level credit assignment in long-horizon video reasoning, VISD introduces a video-aware judge guided self-distillation framework with direction–magnitude decoupling, where verifiable rewards determine update directions while structured privileged signals modulate token-level magnitudes, achieving higher accuracy, stronger grounding, and nearly \(2\times\) faster convergence.
- Visual Enhanced Depth Scaling for Multimodal Latent Reasoning
-
To tackle visual under-optimization and fixed-depth bottlenecks in multimodal latent reasoning, this paper introduces Spatially-Coherent Finer Visual Replay (SCF-VR) and Routing Depth Scaling (RDS), integrated with curriculum learning for efficient and accurate latent reasoning.
- Visual Latents Know More Than They Say: Unsilencing Latent Reasoning in MLLMs
-
Addressing the optimization pathology where the autoregressive objective bypasses continuous visual latents via direct visual shortcuts—termed Silenced Visual Latents—this paper proposes a training-free, test-time framework that optimizes latent representations under frozen backbones via query-guided contrastive warm-up and confidence-progression reinforcement.
- VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing
-
To overcome the severe perception bias in existing remote sensing vision-language benchmarks, VLRS-Bench establishes the first dedicated complex reasoning benchmark structured across Cognition, Decision, and Prediction, spanning 14 subtasks and 2,000 expert-verified QA pairs.
- What MLLMs Learn about When they Learn about Multimodal Reasoning
-
Introduces MathLens, a formally grounded geometry reasoning benchmark that operationally decomposes multimodal reasoning into perception, text-only reasoning, and multimodal-specific coordination, showing that reinforcement learning primarily drives perceptual grounding while textual SFT indirectly benefits perception through reflective reasoning.
- What’s Holding Back Latent Visual Reasoning?
-
This paper systematically assesses continuous latent visual reasoning in vision-language models, uncovering through causal interventions and geometric probes that models routinely bypass latent tokens due to trivial training data and suffer from severe representation collapse during inference.
- When and How Much to Imagine: Adaptive Test-Time Scaling with World Models for Visual Spatial Reasoning
-
Addressing the computational waste and misleading hallucination noise caused by always-on visual world model exploration in multimodal spatial reasoning, this paper proposes AVIC, an adaptive test-time scaling framework, and its RL-trained counterpart AVIC-R, which selectively decides when and how much to imagine based on visual evidence sufficiency, matching or outperforming always-on baselines and proprietary models with substantially fewer world model calls and token overhead.