AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Does Reasoning Improve Seeing? Understanding When Vision-Language Models Benefit from Thinking

Jing Bi, Luchuan Song, Dingxin Zhang, Pinxin Liu, Guangyu Sun, Lianggong Bruce Wen, Weidong Cai, Chen Chen, Chenliang Xu

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.

Investigates whether VLMs encode meta-cognitive signals indicating when explicit reasoning (Thinking mode) helps versus direct Instruct answering, to enable adaptive test-time inference.

Problem

Practitioners lack principled ways to decide when reasoning improves vision-language model performance and how much test-time computation to spend.

Approach

Derives oracle labels for reasoning helpfulness (whether Thinking mode beats Instruct mode on a sample) and probes whether VLM internal representations encode these signals for adaptive inference.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use Thinking-model representations instead of Instruct for linear probing of reasoning helpfulness (ablation)linear probe on final-layer last-token embedding · MMMU-val900, InternVL-8BMSE (r_help): 0.168 -> 0.119Improved
use Thinking-model representations instead of Instruct for probing desired length (ablation)linear probe on last-token embedding · MMMU-val900, Qwen3-VL-32BMAE (l_opt): 39.1 -> 31.8Improved
increase model size from 8B to 32B (ablation)Thinking-probe for r_help · MMMU-val900, Qwen3-VLSpearman rho: 0.447 -> 0.531Improved
scale length-control heads (alpha 0.5 to 2.0) (ablation)Thinking VLMs · MMMU, InternVL-8B-Thinkingmean generation length (tokens): 139 -> 521Improved
scale length-control heads (alpha 0.5 to 2.0) (ablation)Thinking VLMs · MMMU, InternVL-8B-Thinkingaccuracy: -0.8%No clear effect
amplify low-layer reasoning heads while suppressing perception heads (ablation)InternVL-8B-Thinking · MMMU perception-dominant cohortaccuracy change: +4.1%Improved
amplify high-layer reasoning heads (ablation)InternVL-8B-Thinking · MMMU reasoning-dominant cohortaccuracy change: +5.3%Improved
depth-controlled reasoning head steering (ablation)InternVL-8B-Thinking · MMMU overalloverall accuracy change: +1.2%Improved
head-scaling steering applied to other benchmarks (ablation)InternVL-8B-Thinking · MathVista, ScienceQA, ChartQA, A-OKVQAaccuracy change: +2.9% to +4.9%Improved
increase number of stochastic passes K from 5 to 10 for labels (ablation)oracle label estimation · Qwen3-VL-4Bcorr(r_K, r_40): 0.885 -> 0.962Improved
vary head-set size K from 8 to 20 (ablation)head attribution · MMMU, averaged across modelsJaccard overlap: 0.71 -> 0.68 for J(Sr,Sd)No clear effect

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