๐ง VLM Reasoning¶
๐ฌ EMNLP2026 ยท 17 paper notes
๐ Same area in other venues: ๐ง NeurIPS2026 (81) ยท ๐๏ธ ECCV2026 (164) ยท ๐ท CVPR2026 (150) ยท ๐ฌ ICLR2026 (112) ยท ๐ฌ ACL2026 (32) ยท ๐งช ICML2026 (31)
๐ฅ Top topics: Reasoning ร13 ยท Multimodal/VLM ร9 ยท Medical Imaging ร2 ยท LLM ร2
- Ada-DeVL: Adaptive Decoupling of Visual and Logical Reasoning Pathways for Efficient VLM Inference
-
Ada-DeVL decouples multimodal test-time reasoning into an orthogonal \(2 \times 2\) taxonomy of visual perception depth and logical reasoning intensity, optimizing dynamic pathway routing via self-distillation SFT and GRPO with diversity/length penalties to drastically curtail token waste and overthinking.
- Decompose, Look, and Reason: Reinforced Latent Reasoning for VLMs
-
The authors introduce DLR, a reinforced latent reasoning framework that dynamically interleaves textual premise generation, premise-conditioned continuous visual latent extraction, and grounded rationale deduction, coupled with a Spherical Gaussian Latent Policy (SGLP) under GRPO that unlocks active latent exploration.
- FloCA: Towards Faithful and Logically Consistent Flowchart Reasoning
-
To tackle the severe hallucination and topology-inconsistent transitions in flowchart-oriented dialogue, FloCA decouples language understanding from graph execution via an external topology-constrained reasoning tool, achieving state-of-the-art task success and path coverage.
- How Robust is OCR-Reasoning? Evaluating OCR-Reasoning Robustness of Vision-Language Models under Visual Perturbations
-
This paper introduces OCR-Robust, a controlled diagnostic benchmark evaluating vision-language models on OCR reasoning under visual perturbations, demonstrating that higher reasoning capability and larger scale do not guarantee robustness, and that structured visual representations such as charts and tables suffer pronounced worst-case degradation.
- Keeping Eyes on the Image: Dynamic Visual Attention Recalibration for Vision-Language Reasoning
-
To tackle visual attention decay in Large Vision-Language Models (LVLMs) during long-form Chain-of-Thought reasoningโwhere models progressively ignore image evidence in favor of generated textโthis paper proposes DVAR (Dynamic Visual Attention Recalibration), a training-free inference-time method that dynamically amplifies visual token value vectors via spatial semantic weighting and temporal shift scheduling, consistently boosting reasoning fidelity across six benchmarks.
- Learning to Reason Before Masking for Referring Audio-Visual Segmentation
-
To address the lack of multi-hop cognitive reasoning over implicit clues in referring audio-visual segmentation, this paper introduces the Reasoning Ref-AVS task, the ReAVS benchmark, and Re2Seg, an omnimodal framework that reasons over candidate-entity evidence before conditioning mask decoding.
- MedQA-MM: Shortcuts Behind Medical Visual Reasoning
-
Addressing the "reasoning inflation" crisis where medical multimodal multiple-choice benchmarks conflate surface accuracy with clinical visual reasoning, this paper establishes an evidence-ladder auditing framework with matched minimal repairs across six benchmarks, and releases MEDQA-MM, a 1,000-item shortcut-mitigated evaluation benchmark.
- MeetingToM: Evaluating Multimodal LLMs on Theory-of-Mind Reasoning in Multi-Party Meetings
-
Addressing the gap where existing Theory-of-Mind evaluations focus on overt, externally verifiable video QA, MeetingToM establishes a hierarchical benchmark spanning individual mental states, dyadic addressee-attitude pairs, and group pseudo-consensus across 60 naturalistic multi-party meetings, revealing that leading MLLMs struggle with multimodal cue calibration and suffer severe bias when distinguishing genuine agreement from concealed dissent.
- MemDreamer: Decoupling Perception and Reasoning for Long Video Understanding via Dreaming-Induced Dynamic Working Memory
-
MemDreamer decouples long video understanding into streaming perceptual memory construction and active agentic reasoning over a pure-text hierarchical graph, achieving state-of-the-art results (85.5 on LVBench) while cutting the active reasoning context window down to 5.9Kโ6.3K tokens.
- MR^2-Bench: Going Beyond Matching to Reasoning in Multimodal Retrieval
-
MR^2-Bench introduces the first reasoning-intensive multimodal retrieval benchmark spanning 12 diverse sub-tasks across knowledge retrieval, illustration search, and visual relation reasoning, revealing that SOTA embedding models collapse from ~78% on standard benchmarks to under 10% Recall@1, while test-time CoT rewriting and multimodal reasoning rerankers unlock substantial headroom up to +15.22 nDCG@10.
- PIVOTS-Bench: Evaluating Fine-Grained Interpersonal Relationship Reasoning in Multimodal Large Language Models
-
PIVOTS-Bench introduces the first multimodal benchmark for evaluating fine-grained bidirectional interpersonal relationship reasoning across six psychological dimensions, revealing substantial deficiencies of current MLLMs in grounded visual causality and implicit social dynamics.
- PIVOTS-Bench: Evaluating Fine-Grained Interpersonal Relationship Reasoning in Multimodal Large Language Models
-
This paper introduces PIVOTS-Bench, the first multimodal benchmark evaluating fine-grained interpersonal relationship reasoning in MLLMs across six bidirectional psychological dimensions, revealing pronounced performance gaps and unexpected behavioral reversals between explicit dialogues and nuanced social interactions.
- ReactBench: A Benchmark for Topological Reasoning in MLLMs on Chemical Reaction Diagrams
-
ReactBench introduces the first benchmark diagnosing topological reasoning in MLLMs on chemical reaction diagrams via 1,618 expert-annotated QA pairs across four hierarchical dimensions, revealing that while frontier models excel at local anchor extraction (>90%), global structural reasoning suffers a persistent drop (>30%) driven by reasoning bottlenecks rather than perception failures.
- Think, Act, Build: An Agentic Framework with Vision Language Models for Zero-Shot 3D Visual Grounding
-
TAB reformulates zero-shot 3D visual grounding into a dynamic "Think, Act, Build" agentic paradigm over continuous RGB-D streams, completely eliminating dependence on preprocessed 3D point clouds and offline 3D detectors through an innovative Semantic-Anchored Geometric Expansion mechanism that actively acquires multi-view masks via deterministic projection.
- Think, Look, and Revise: Inconsistency-Aware Visual Self-Correction in MLLMs
-
Addressing the open-loop limitation of tool-augmented multimodal large language models (MLLMs) that blindly trust intermediate perceptual outputs, this paper proposes ReVISE, a closed-loop framework combining reflection-supervised cold-start trajectories with correction- and grounding-oriented reinforcement learning to enable active evidence verification, dynamic re-querying, and fallback to native perception.
- VITAL: Visual-Semantic Dual Supervision for Enhanced and Interpretable Latent Reasoning in Medical MLLMs
-
VITAL proposes a latent-space reasoning framework for medical MLLMs that enforces strict train-inference parity through a recurrent loop and prevents latent state modality collapse using orthogonal training-time scaffolding (an auxiliary text decoder and an independent MedSigLIP visual projector), achieving comprehensive SOTA across 7 benchmarks with up to 97x speedup over explicit CoT and zero-overhead post-hoc dual interpretability.
- Which one is banana man? Evaluating visionโlanguage models in multi-turn pragmatic interpretation
-
By benchmarking leading visionโlanguage models on transcripts of human iterated tangram reference games across 8 counterfactual context conditions and 3 feedback levels, this paper demonstrates that VLMs resolve referring expressions largely through shallow semantic retrieval rather than dynamic pragmatic adaptation, failing to self-correct past errors and struggling severely with metaphorical conventions.