🧠 VLM Reasoning¶

💬 ACL2026 · 32 paper notes

📌 Same area in other venues: 📷 CVPR2026 (150) · 🔬 ICLR2026 (112) · 🧪 ICML2026 (31) · 🤖 AAAI2026 (10) · 🧠 NeurIPS2025 (30) · 📹 ICCV2025 (15)

🔥 Top topics: Reasoning ×30 · Multimodal/VLM ×19 · LLM ×4 · Agents ×2

A Survey of Multimodal Mathematical Reasoning: From Perception, Alignment to Reasoning: This survey proposes a complementary perspective consisting of the Perception–Alignment–Reasoning (PAR) process framework and the Answer–Process–Executable (APE) evaluation framework. It systematically organizes three task families—geometry, chart/table, and visual word problems—mapping existing methods and benchmarks onto these two coordinate axes. It represents the first process-centric survey on multimodal mathematical reasoning.
Addressing Overthinking in Large Vision-Language Models via Gated Perception-Reasoning Optimization: The GPRO framework is proposed to address overthinking in LVLMs by dynamically routing computation to three paths (Fast/Perception Re-check/Reasoning Reflection) at each token generation step through a meta-reasoning controller, simultaneously improving both accuracy and efficiency.
AnchorSeg: Language Grounded Query Banks for Reasoning Segmentation: Ours proposes AnchorSeg, reframing reasoning segmentation as a structured conditional generation process based on a language-grounded query bank. It explicitly decouples spatial localization and semantic reasoning via anchor queries and a Token-Mask cycle consistency training objective, achieving SOTA on ReasonSeg (67.7% gIoU, 68.1% cIoU).
ArrowGEV: Grounding Events in Video via Learning the Arrow of Time: ArrowGEV is proposed, a reinforcement learning framework inspired by the "Arrow of Time" in physics. It models temporal directionality by distinguishing between time-sensitive and time-insensitive events, enhancing the event grounding accuracy and temporal understanding of VLMs.
Can MLLMs Reason Beyond Language? VisReason: A Comprehensive Benchmark for Vision-Centric Reasoning: VisReason constructs a multimodal benchmark containing 1,505 daily visual reasoning problems to specifically test whether models can reason directly based on visual evidence. Results show that even the strongest model achieves an average accuracy of only 47.5%, significantly lower than the human performance of 71.4%, and that CoT and larger reasoning budgets provide limited improvements.
CAPruner: Conceptual-Adjacent Scene Graph Pruner for Enhancing 3D Spatial Reasoning of Large Language Models: To address the conflict where feeding complete 3D scene graphs to LLMs leads to token explosion while existing distance-based KNN pruning often removes task-critical relations, this paper proposes CAPruner. It integrates "query semantic relevance" and "spatial proximity" into a lightweight MLP (only 1219 parameters) to score the importance of each edge. The model is trained via weak supervision by "aggregating edge weights into node weights" using only target object labels. Under a fixed edge budget, it preserves relations truly useful for specific 3D-VL tasks, significantly improving the spatial reasoning accuracy of downstream LLMs.
ChemVLR: Prioritizing Reasoning in Perception for Chemical Vision-Language Understanding: ChemVLR is proposed as the first reasoning-based VLM in the chemical domain. It constructs a 760K reasoning dataset via a cross-modal reverse engineering strategy and employs a three-stage training pipeline (CPT-SFT-RL), significantly outperforming proprietary models and domain-specific VLMs in molecular recognition and reaction prediction.
Decoding Scientific Experimental Images: The SPUR Benchmark for Perception, Understanding, and Reasoning: SPUR is the first benchmark designed for the "Perception \(\rightarrow\) Understanding \(\rightarrow\) Reasoning" three-stage evaluation of biomedical experimental images (multi-panel staining, Western blots, and statistical charts). It contains 4,264 expert-verified MCQs, revealing that current MLLMs (with Gemini 3 Pro Preview barely exceeding 60%) generally perform 12.76%–31.41% lower in quantitative reasoning than in qualitative reasoning.
Do MLLMs Understand Pointing? Benchmarking and Enhancing Referential Reasoning in Egocentric Vision: The authors construct EgoPoint-Bench, the first hybrid real+physical simulation benchmark for egocentric "finger pointing" QA (11.7k QA / 5 dimensions / 3 semantic referential levels). They confirm that current SOTA MLLMs generally rely on "visual proximity / saliency" pseudo-correlations rather than truly parsing the fingertip ray. Through LoRA fine-tuning on simulated data, they achieve an average improvement of up to +25 points and robust sim-to-real generalization.
DRIFT: Transferring Reasoning Priors for Efficient MLLM Fine-Tuning: DRIFT treats the "parameter difference between a text reasoning expert and a multimodal model" as a directional prior. During multimodal SFT backpropagation, it applies a lightweight bias to gradients (without modifying weights). Using only 4K multimodal CoT data and approximately 2 hours of training, it consistently pushes Qwen2.5-VL-7B performance on benchmarks like MathVista, MathVerse, and WeMath beyond parameter merging baselines and heavy SFT/RL methods.
ErrorRadar: Benchmarking Complex Mathematical Reasoning of Multimodal Large Language Models Via Error Detection: This paper formally defines the multimodal error detection task and constructs the ErrorRadar benchmark—comprising 2,500 K-12 multimodal math problems derived from real student responses. It evaluates MLLM capabilities in two subtasks: error step localization (STEP) and error type classification (CATE), finding that the strongest model, GPT-4o, still trails human evaluation by approximately 10-15%.
Faithful-First Reasoning, Planning, and Acting for Multimodal LLMs: This paper proposes the Faithful-First RPA framework, which evaluates perceptual faithfulness (whether claimed objects truly exist in the image) at each reasoning step via the FaithEvi pipeline and enforces evidence-based planning and action through the FaithAct mechanism during generation. This approach improves perceptual faithfulness by up to 24% without compromising task accuracy.
Forest Before Trees: Latent Superposition for Efficient Visual Reasoning: This paper proposes Laser, which performs visual reasoning in latent space via Dynamic Window Alignment Learning (DWAL). By maintaining a "probabilistic superposition" of future semantics rather than precise token-by-token prediction, the model achieves a "global-to-local" cognitive hierarchy. Laser reaches SOTA among latent reasoning methods across six benchmarks using only 6 reasoning tokens (a 97%+ reduction), outperforming Monet by an average of 5.03%.
GeoArena: Evaluating Open-World Geographic Reasoning in Large Vision-Language Models: Ours proposes GeoArena, a "dynamic, label-free, process-oriented" evaluation platform for open-world geographic reasoning in LVLMs. It reformulates geographic localization assessment under in-the-wild images into a pairwise reasoning alignment task, ranking 17 frontier LVLMs using human preferences and Bradley-Terry scores, achieving a 78% expert-crowdsource agreement rate.
GeoRC: A Benchmark for Geolocation Reasoning Chains: Proposes GeoRC, the first geolocation reasoning chain benchmark written by GeoGuessr champion-level experts (800 reasoning chains, 500 scenes). It evaluates the ability of VLMs to generate auditable reasoning chains, finding that while closed-source VLMs can match human localization accuracy, the quality of their reasoning chains lags significantly, while open-source VLMs are almost equivalent to a pure hallucination baseline.
HierVA: Hierarchical Visual Agent — Managing Contexts in Joint Image-Text Space for Advanced Chart Reasoning: HierVA utilizes a "manager–worker" dual-layer multimodal agent to manage both image and text contexts during chart reasoning through a disciplined "acquisition–limitation–distillation" approach. It outperforms strong baselines like CoT and "thinking with images" on complex chart reasoning benchmarks such as CharXiv in a training-free manner.
iReasoner: Trajectory-Aware Intrinsic Reasoning Supervision for Self-Evolving Large Multimodal Models: iReasoner enables LMMs to perform self-questioning and answering on unlabeled images, extending final answer consistency into an intrinsic reward for intermediate CoT steps. This leads to a multimodal reasoning improvement of up to +2.13 points on Qwen2.5-VL-7B.
MMErroR: A Benchmark for Erroneous Reasoning in Vision-Language Models: This paper proposes MMErroR, a multi-modal erroneous reasoning benchmark containing 1,997 samples. Each sample embeds a single reasoning error across 6 domains and 4 error types. It requires VLMs to not only detect the presence of errors in a reasoning chain but also classify the error type (Vision Perception, Knowledge Application, Question Understanding, or Reasoning Error). Evaluation of 12 representative VLMs reveals that the strongest model, Gemini-3-Pro-Preview, achieves only 66.65% accuracy.
OMHBench: Benchmarking Balanced and Grounded Omni-Modal Multi-Hop Reasoning: OMHBench constructs an omni-modal three-hop reasoning benchmark with 6,144 questions covering text, image, and speech contexts. Through entity-attribute chains and 6 balanced reasoning paths, it exposes systematic shortcomings of current MLLMs in speech grounding, path robustness, and cross-modal grounding.
OMIBench: Benchmarking Olympiad-Level Multi-Image Reasoning in Large Vision-Language Models: This paper introduces OMIBench—the first large-scale benchmark for Olympiad-level multi-image reasoning. It covers over 1000 competition problems across biology, chemistry, mathematics, and physics. The study finds that even the strongest LVLM (Gemini-3-Pro) achieves only about 50% accuracy, representing a drop of over 25% compared to single-image benchmarks.
Position: Multimodal Large Language Models Can Significantly Advance Scientific Reasoning: This is a position paper advocating that Multimodal Large Language Models (MLLMs) can significantly advance interdisciplinary scientific reasoning. It proposes a four-stage research roadmap (Broad Knowledge Recognition → Analogical Generalization → Insightful Reasoning → Creative Hypothesis Generation) and provides a systematic review of MLLM applications across mathematics, physics, chemistry, and biology, identifying five major challenges and eight future directions.
PROGRESSLM: Towards Progress Reasoning in Vision-Language Models: This paper defines the ability to "judge task completion stages from a single-frame observation" as the VLM's progress reasoning capability. It constructs Progress-Bench and ProgressLM-45K, demonstrating that explicit learning of "episodic retrieval + mental simulation" is more stable than simple prompt-based reasoning.
SciMDR: Advancing Scientific Multimodal Document Reasoning: SciMDR proposes a synthesize-and-reground data construction framework that first synthesizes faithful QA and reasoning chains based on atomic claims, then re-embeds them into full scientific papers for model training. This enables a 7B VLM to approach GPT-5 series performance in scientific multimodal document reasoning.
ShredBench: Evaluating the Semantic Reasoning Capabilities of Multimodal LLMs in Document Reconstruction: ShredBench constructs an evaluation benchmark centered on "restoring content from shredded documents using multimodal large language models," demonstrating that current MLLMs, despite being strong in conventional OCR, generally lack the ability to perform reasoning by integrating visual fragments, reading order, and semantic context.
Structured and Abstractive Reasoning on Multi-modal Relational Knowledge Images: This paper proposes the STAR data engine and a two-stage training framework for multi-modal relational knowledge images. By utilizing STAR-64K synthetic data, Chain-of-Thought (CoT) annotations, and knowledge-aware KGRPO, it significantly enhances the capability of Multimodal Large Language Models (MLLMs) to understand and reason over abstract structured knowledge images.
TableVista: Benchmarking Multimodal Table Reasoning under Visual and Structural Complexity: TableVista constructs a multimodal table benchmark with 3,000 high-quality reasoning questions expanded into 30,000 visual samples. Systematic evaluation of 29 foundation models reveals that while models are relatively stable across style changes, they suffer significant degradation under complex structures, multi-table reasoning, visual fragmentation, and vision-only inputs.
TemporalVLM: Video LLMs for Temporal Reasoning in Long Videos: This paper proposes TemporalVLM, which extracts local fine-grained temporal features through a time-aware segment encoder (overlapping sliding Video Q-Former + fusion module) and aggregates global long-range dependencies using a BiLSTM. This marks the first introduction of LSTM into Video LLMs, outperforming previous methods across four tasks: dense video captioning, temporal localization, highlight detection, and action segmentation.
Thinking Like a Botanist: Challenging Multimodal Language Models with Intent-Driven Chain-of-Inquiry: This paper introduces the PlantInquiryVQA benchmark and the Chain-of-Inquiry (CoI) framework, comprising 24,950 plant images and 138,068 QA pairs. It simulates the adaptive diagnostic questioning strategies of botanists to evaluate the multi-step visual reasoning capabilities of 18 MLLMs in plant pathology. Findings show that structured questioning significantly improves diagnostic accuracy and reduces hallucinations, though even the strongest models achieve a clinical utility score of only 0.188.
TRACE: Unleashing Spatial Reasoning in Multimodal Large Language Models via Textual Representation Guided Reasoning: This paper proposes TRACE (Textual Representation of Allocentric Context from Egocentric Video), a prompting method that guides Multimodal Large Language Models (MLLMs) to generate structured textual allocentric 3D environment representations—including meta-context, camera trajectories, and entity registries—from egocentric videos. These serve as intermediate reasoning steps to enhance spatial question-answering capabilities, consistently outperforming existing prompting strategies on VSI-Bench and OST-Bench.
VL-Calibration: Decoupled Confidence Calibration for Large Vision-Language Models Reasoning: VL-Calibration decouples the verbalized confidence of LVLMs into visual confidence and reasoning confidence. By utilizing image-perturbation KL divergence, token entropy, and token-level advantage reweighting for training, the model simultaneously reduces ECE and improves accuracy across 13 visual reasoning benchmarks.
What's Missing in Screen-to-Action? Towards a UI-in-the-Loop Paradigm for Multimodal GUI Reasoning: This paper proposes the UILoop (UI-in-the-Loop) paradigm, reframing GUI reasoning from the traditional "Screen → Action" into a "Screen → UI Element → Action" cyclic process. Through UI element-driven reinforcement fine-tuning, the model is taught to explicitly locate, understand, and utilize key UI elements, achieving SOTA performance on GUI reasoning tasks.
When Slower Isn't Truer: Inverse Scaling Law of Truthfulness in Multimodal Reasoning: This paper identifies an "Inverse Scaling Law" in multimodal reasoning—reasoning (slow-thinking) models are more prone to generating untruthful outputs than chat (fast-thinking) models when facing misleading visual inputs. To systematically diagnose this, the authors constructed the TruthfulVQA benchmark (5,000+ samples, 50 annotators, three-level hierarchical prompts) and the TruthfulJudge evaluation model (88.4% accuracy).