AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Awakening Visual Reasoning: Mitigating Post-Training Failure in Vision-Text Compression

Xing Xi, Liyao Li, Hao Chen, Ningtao Wang, Peixian Chen, Peilin Tong, Xing Fu, Yu Cheng, Haobo Wang, Gang Chen, Junbo Zhao, Ronghua Luo

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis

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.

Studies why vision-text compression (rendering text as dense visual tokens) degrades reasoning in VLMs even though OCR fidelity stays high, and proposes a mitigation for this post-training failure.

Problem

Vision-text compression renders long text into dense visual tokens for scalable long-context multimodal modeling, but VLMs that reason well on native text fail on visually compressed equivalents, especially in long-range retrieval and multi-step reasoning.

Approach

Identifies a reasoning gap between native-text and visually compressed inputs despite high OCR fidelity, and targets post-training as the source of the failure; the abstract does not specify the method in detail.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add CoRe alignment to RL on visual promptsVisual GRPO · Math average (visual)accuracy (%): 29.44 -> 42.97Improved
add CoRe alignment to SFT on visual promptsVisual SFT · Math average (visual)accuracy (%): 15.42 -> 20.92Improved
add CoRe alignment to RLVisual GRPO · GSM8K (visual)accuracy (%): 67.31 -> 82.19Improved
add CoRe alignment to RLVisual GRPO · HotpotQA (visual)accuracy (%): 9.00 -> 23.00Improved
add CoRe alignment to RLVisual GRPO · 2WikiMultihopQA (visual)accuracy (%): 13.00 -> 35.00Improved
add CoRe alignment to RLVisual GRPO · InfoTabs (visual)accuracy (%): 61.29 -> 70.18Improved
add CoRe alignment to RLVisual GRPO · HiTab (visual)accuracy (%): 3.16 -> 5.06Improved
add CoRe alignment to RLVisual GRPO · Qasper (visual)accuracy (%): 1.00 -> 4.00Improved
apply standard GRPO to visual prompts instead of textVisual GRPO · HotpotQAaccuracy (%): 7.0% to 9.0%Improved
use ground-truth CoT alignment (CoRe-SFT) rather than one self-generated rollout (K = 1) (ablation)CoRe · AIME 2024accuracy (%): 3.33 -> 6.67Improved
use self-generated rollout (K = 1) rather than ground-truth trace (ablation)CoRe · MATH-500accuracy (%): 42.20 -> 42.80No clear effect
increase rollouts from K = 1 to K = 4 (ablation)CoRe-RL · MATH-500accuracy (%): 42.80 -> 47.20Improved

Surprising findings

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