AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

DocThinker: Explainable Multimodal Large Language Models with Rule-based Reinforcement Learning for Document Understanding

Wenwen Yu, Zhibo Yang, Yuliang Liu, Xiang Bai

ICCV 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 11 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace SFT with GRPO RL training (336 input, 4k)Qwen2.5-VL-7B · DocVQAscore: 0.355 -> 0.579Improved
replace SFT with GRPO RL training (336 input, 4k)Qwen2.5-VL-7B · TextVQAscore: 0.740 -> 0.802Improved
replace SFT with GRPO RL training (336 input, 4k)Qwen2.5-VL-7B · GQAscore: 0.467 -> 0.546Improved
replace SFT with GRPO RL training (336 input, 4k)Qwen2.5-VL-7B · Visual7Wscore: 0.563 -> 0.580Improved
add RoI IoU reward (ablation)DocThinker-7B · InfoQAscore: 0.637 -> 0.689Improved
add RoI IoU reward (ablation)DocThinker-7B · TextCapsscore: 0.693 -> 0.738Improved
add rephrase question reward (ablation)DocThinker-7B · TextVQAscore: 0.772 -> 0.827Improved
add rephrase question reward (ablation)DocThinker-7B · DocVQAscore: 0.763 -> 0.795Improved
add both RoI IoU and rephrase question rewards (ablation)DocThinker-7B · InfoQAscore: 0.602 -> 0.689Improved
add KL regularization beta=0.04 versus none (ablation)DocThinker-7B · TextVQAscore: 0.803 -> 0.827Improved
increase KL beta from 0.001 to 0.04 (ablation)DocThinker-7B · DocVQAscore: 0.785 -> 0.795Improved
increase input resolution from 336 to 1536DocThinker-7B · DocVQAscore: 0.579 -> 0.795Improved

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