# MMSkills: Towards Multimodal Skills for General Visual Agents

**Authors:** Kangning Zhang, Shuai Shao, Qingyao Li, Jianghao Lin, Lingyue Fu, Shijian Wang, Wenxiang Jiao, Yuan Lu, Weiwen Liu, Weinan Zhang, Yong Yu  
**arXiv:** [2605.13527](https://arxiv.org/abs/2605.13527)  
**Date:** 2026-05-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2605.13527v1/

## Key points

- Reusable skills are crucial for improving agent capabilities
- MMSkills framework introduced to represent, generate, and utilize reusable multimodal procedures for real-time visual decision-making
- Each MMSkill combines textual procedure with runtime state cards and multi-view keyframes
- Experiments show that MMSkills enhance performance of both frontier and smaller multimodal agents in visual decision-making tasks

## Summary

, , , , 

In the realm of improving agent capabilities, reusable skills have emerged as a fundamental component. However, traditional skill packages often encode reusable behavior in the form of textual prompts, executable code, or learned routines. To address this need for multimodal procedural knowledge, we introduce MMSkills—a framework designed to represent, generate, and utilize reusable multimodal procedures for real-time visual decision-making. Each MMSkill encapsulates a compact package that combines a textual procedure with runtime state cards and multi-view keyframes. Our experiments across GUI and game-based visual-agent benchmarks demonstrate that MMSkills consistently enhance both frontier and smaller multimodal agents. This suggests that external multimodal procedural knowledge complements internal model priors effectively. Overall, our work emphasizes the importance of incorporating multimodal procedural knowledge into agent capabilities and showcases how MMSkills can significantly improve performance in visual decision-making tasks.