Research map / Vision-language and multimodal / Multimodal LLMs
Bridging Visual Representation and Reinforcement Learning from Verifiable Rewards in Large Vision-Language Models
yuhang han, Yuyang Wu, Zhengbo Jiao, Yiyu Wang, Xuyang Liu, Shaobo Wang, Hanlin xu, Xuming Hu, Linfeng Zhang
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add KAWHI to GRPO | GRPO · Math reasoning avg (MathVista, MathVerse, MathVision, WeMath), Qwen2.5-VL-7B | average accuracy: 50.18 -> 51.00 | Improved |
| add KAWHI to DAPO | DAPO · Math reasoning avg (MathVista, MathVerse, MathVision, WeMath), Qwen2.5-VL-7B | average accuracy: 50.79 -> 51.85 | Improved |
| add KAWHI to GSPO | GSPO · Math reasoning avg (MathVista, MathVerse, MathVision, WeMath), Qwen2.5-VL-7B | average accuracy: 51.40 -> 52.94 | Improved |
| add KAWHI to GRPO | GRPO · Math reasoning avg (MathVista, MathVerse, MathVision, WeMath), Qwen3-VL-4B | average accuracy: 51.17 -> 51.98 | Improved |
| add KAWHI to DAPO | DAPO · Math reasoning avg (MathVista, MathVerse, MathVision, WeMath), Qwen3-VL-4B | average accuracy: 51.80 -> 53.06 | Improved |
| add KAWHI to GSPO | GSPO · Math reasoning avg (MathVista, MathVerse, MathVision, WeMath), Qwen3-VL-4B | average accuracy: 54.14 -> 55.54 | Improved |
| add KAWHI to GRPO | GRPO · MathVista, Qwen2.5-VL-7B | accuracy: 68.37 -> 69.30 | Mixed |
| add KAWHI to GRPO (chart training) | GRPO · CharXiv Desc | accuracy: 70.1 -> 72.2 | Improved |
| add KAWHI to GRPO (chart training) | GRPO · ChartMimic | accuracy: 42.8 -> 45.1 | Improved |
| use SGUF-selected regions instead of all vision tokens (ablation) | GRPO+KAWHI · MathVerse | accuracy: 46.58 -> 47.56 | Improved |
| use SGUF-selected regions instead of inverse selection (ablation) | GRPO+KAWHI · MathVerse | accuracy: 45.74 -> 47.56 | Improved |
| use key-query instead of key-key similarity (ablation) | GRPO+KAWHI · MathVerse | accuracy: 47.08 -> 47.56 | Improved |
Closest papers by meaning
- RewardMap: Tackling Sparse Rewards in Fine-grained Visual Reasoning via Multi-Stage Reinforcement Learning (ICLR 2026 · similarity 0.91)
- SSL4RL: Revisiting Self-supervised Learning as Intrinsic Reward for Visual-Language Reasoning (ICML 2026 · similarity 0.91)
- Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.91)
- PDCR: Perception-Decomposed Confidence Reward for Vision-Language Reasoning (CVPR 2026 · similarity 0.90)
- Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.90)
- Unveiling the Compositional Ability Gap in Vision-Language Reasoning Model (NeurIPS 2025 · similarity 0.90)
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.
