AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Context-Aware Reasoner: Enhancing Contextual Reasoning in Multimodal Large Language Models

Zhe Zheng, Wenqi Zhang, Xiaohe Zhou, Guiyang Hou, Xinyu Mei, Tianyang Mao, Anni PIAO, Shengpei Jiang, Yongliang Shen, Yueting Zhuang, Weiming Lu

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces ContextReasonV-Bench to evaluate contextual reasoning in multimodal LLMs, identifying two failure modes, and proposes Context-Aware Reasoner to address them (details beyond the abstract unavailable).

Problem

MLLMs are evaluated mainly on reasoning over internalized knowledge, while reasoning grounded in information present in the context (contextual reasoning) is overlooked.

Approach

Constructs a benchmark specifically for visual contextual reasoning and uses its failure analysis to motivate a context-aware reasoning approach; the specific method is not described in the available abstract.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add context-aware reward to vanilla GRPO (ablation)CAR RL stage · ContextReasonV-Bench Overallaccuracy (%): 33.20 -> 34.59Improved
add hierarchical advantage estimation with context augmentation (ablation)CAR RL stage · ContextReasonV-Bench Overallaccuracy (%): 34.59 -> 38.14Improved
add context semantic augmentation (without HAE) (ablation)CAR RL stage · ContextReasonV-Bench Overallaccuracy (%): 34.59 -> 36.74Improved
add HAE advantage estimation on top of augmented contexts (ablation)CAR RL stage · ContextReasonV-Bench Overallaccuracy (%): 36.74 -> 38.14Improved
add SFT on 388 analyze-then-solve samplesQwen2.5-VL-7B · ContextReasonV-Bench Overallaccuracy (%): 16.05 -> 27.68Improved
add RL stage after SFTCAR-SFT · ContextReasonV-Bench Overallaccuracy (%): 27.68 -> 38.14Improved
SFT on out-of-distribution format taskCAR-SFT · Visual Logical Transformationaccuracy (%): 25.40 -> 23.01Hurt
SFT on out-of-distribution format taskCAR-SFT · Symbolic Transformation Learningaccuracy (%): 35.00 -> 15.00Hurt
rule-based math RL (Vision-R1-7B) on base modelQwen2.5-VL-7B · ContextReasonV-Bench Overallaccuracy (%): 16.05 -> 14.85No clear effect
rule-based math RL (MM-Eureka-7B) on base modelQwen2.5-VL-7B · ContextReasonV-Bench Overallaccuracy (%): 16.05 -> 16.38No clear effect
apply CAR trainingQwen2.5-VL-7B · General vision tasks averageaccuracy (%): 45.58 -> 49.60Improved
apply CAR trainingQwen2.5-VL-7B · MMMUaccuracy (%): 53.78 -> 55.56Improved

Limitations

Experimental setup

Benchmarks: ContextReasonV-Bench

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy