---
title: "Visually-Guided Policy Optimization for Multimodal Reasoning"
created: 2026-04-15
updated: 2026-04-15
type: concept
tags: [rl, multimodal, vision, reasoning, policy-optimization]
sources: [raw/papers/2026/04/2604.09349.md]
---
# Visually-Guided Policy Optimization

## 核心方法
视觉引导的策略优化：让 RL 训练的信号同时考虑视觉证据和语言推理。

## 与相关工作的关系
- [[perception-grounded-po]] 的姐妹工作
- 侧重于推理过程中的视觉注意力引导
- 在多模态推理 benchmark 上显著提升

## Related
- [[perception-grounded-po]]
- [[faithful-grpo]]
- [[omni-modal-llm]]