AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models

Yin Zhang, Zonghan Wu, Jiaxuan Zhao, Junfeng Fang, Zengxiang Li, Kun Wang, Qingsong Wen, Yilei Shao

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual 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.

Proposes a mutual-information-guided RLVR method for vision-language models to improve visual grounding before reasoning, based only on the truncated abstract.

Problem

VLMs suffer from visual perception errors and hallucinations that hurt answer accuracy in complex reasoning; existing RLVR methods waste sampling budget and provide weak visual grounding signals.

Approach

Uses mutual information between visual input and responses to guide RL training so the policy attends to visual evidence before reasoning ("see first, reason later"). Details beyond the truncated abstract are not stated.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add MI-guided selection and fork sampling (MIRL-8) vs DAPO-8DAPO · Six-benchmark averageaccuracy: 66.27 -> 68.38Improved
add MI-guided selection (MIRL-12) vs DAPO-16 with fewer trajectoriesDAPO · Six-benchmark averageaccuracy: 69.76 -> 70.22Improved
MIRL-8 with 8 trajectories vs DAPO-12 with 12DAPO · Six-benchmark averageaccuracy: 69.17 -> 68.38No clear effect
add top-MI selection instead of random selection (ablation)MIRL-12 · Six-benchmark averageaccuracy: 68.74 -> 70.22Improved
add top-MI selection instead of bottom-MI selection (ablation)MIRL-12 · Six-benchmark averageaccuracy: 68.02 -> 70.22Improved
add top-MI selection instead of random selection (ablation)MIRL-12 · Math benchmarksaccuracy: 72.58 -> 73.35Improved
add top-MI selection instead of random selection (ablation)MIRL-12 · CoTaccuracy: 70.92 -> 74.35Improved
add top-MI selection instead of random selection (ablation)MIRL-12 · General VLaccuracy: 61.88 -> 63.46Improved
add MI reward on description tokens (lambda 0.1) to task-only reward (ablation)MIRL-12 · Six-benchmark averageaccuracy: 68.37 -> 70.22Improved
use decoupled lambda 0.1 instead of MI-only lambda 1.0 (ablation)MIRL-12 · Six-benchmark averageaccuracy: 68.81 -> 70.22Improved
add MI reward to task-only reward (ablation)MIRL-12 · General VLaccuracy: 65.23 -> 63.46Hurt
MI-only reward vs task-only reward (ablation)MIRL-12 · Six-benchmark averageaccuracy: 68.37 -> 68.81No clear effect

Experimental setup

Builds on: RLVR

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