Add MUSE: A Unified Agentic Harness for MLLMs

[Jianglin954]

Hey, thank you for putting together this valuable collection!

I would appreciate if you could include our recent work:

MUSE: A Unified Agentic Harness for MLLMs

• Paper: https://arxiv.org/abs/2606.03005
• Code: https://github.com/Jianglin954/MUSE

Unlike most existing harness optimization methods that are primarily developed and evaluated on text-centric domains (e.g., coding, match reasoning, and Terminal-Bench-style tasks), MUSE investigates harness design for multimodal large language models (MLLMs).

The paper presents a unified agentic harness and evaluates it across a diverse set of vision-centric tasks, including visual spatial planning, visual perception, multimodal reasoning, and fine-grained visual discrimination.

The results demonstrate that harness optimization can provide substantial gains beyond language-only settings and serves as a general framework for improving multimodal intelligence.

Since this repository already covers harness engineering and agent optimization research, we believe MUSE would be a relevant addition.

Thank you for your consideration.

@article{lu2026museunifiedagenticharness,
      title={MUSE: A Unified Agentic Harness for MLLMs}, 
      author={Jianglin Lu and Hailing Wang and Xu Ma and Qihua Dong and Mingyuan Zhang and Yizhou Wang and Yun Fu},
      year={2026},
      journal={arXiv preprint arXiv:2606.03005},
      url={https://arxiv.org/abs/2606.03005}, 
}