AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Vision-G1: Towards General Reasoning Vision-Language Models via Reinforcement Learning

Yuheng Zha, Kun Zhou, Yujia Wu, Yushu Wang, Jie Feng, Zhi Xu, Shibo Hao, Zhengzhong Liu, Eric P. Xing, Zhiting Hu

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 2 citations (OpenAlex)

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.

Vision-G1 applies reinforcement learning from verifiable rewards to a vision-language model to improve general visual reasoning across diverse domains, addressing the scarcity of verifiable reward data beyond math and logic.

Problem

Current VLMs trained with RLVR cover a narrow range of reasoning tasks (mainly math and logic) because verifiable reward data is scarce in broader domains, so they generalize poorly.

Approach

Uses reinforcement learning from verifiable rewards to train a vision-language model for general reasoning across broader visual domains, targeting the lack of verifiable reward data outside math and logic. Abstract only; method details unavailable.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
RL training with Vision-G1 pipelineQwen2.5-VL-7B · MathVistaaccuracy: 65.0 -> 76.1Improved
RL training with Vision-G1 pipelineQwen2.5-VL-7B · Muir-Benchaccuracy: 39.8 -> 61.5Improved
RL training with Vision-G1 pipelineQwen2.5-VL-7B · VQA-RADaccuracy: 74.5 -> 72.1Hurt
RL training with Vision-G1 pipelineQwen2.5-VL-7B · MMMU-Valaccuracy: 58.6 -> 53.4Hurt
add multi-round RL training (ablation)Vision-G1 · MathVistaaccuracy: 74.9 -> 76.1Improved
add multi-round RL training (ablation)Vision-G1 · ChartQA-Proaccuracy: 48.4 -> 47.7Hurt
add influence-based filtering and difficulty selection (ablation)Vision-G1 · MathVistaaccuracy: 71.5 -> 76.1Improved
add influence-based filtering and difficulty selection (ablation)Vision-G1 · Muir-Benchaccuracy: 59.4 -> 61.5Improved
add domain-specific datasets (ablation)Vision-G1 · MathVistaaccuracy: 76.3 -> 76.1Hurt
add domain-specific datasets (ablation)Vision-G1 · VQA-RADaccuracy: 68.1 -> 72.1Improved
RL training on math-focused data (Vision-R1-7B)Qwen2.5-VL-7B · VQA-RADaccuracy: 74.5 -> 64.9Hurt

Limitations

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