GRSE credits multi-agent skills by contrasting repeated runs
Original title: Which Agent's Skill Should Change? Group-Relative Credit Assignment for Evolving Multi-Agent Teams
On HotpotQA and LoCoMo it beat MASkills on most categories, with the largest gains on multi-hop questions.
The paper presents Group-Relative Skill Evolution (GRSE), which credits agent-skill pairs by running each question several times under the same skill libraries and contrasting those executions. Skill presence alone is treated as insufficient evidence of responsibility. With two unnamed open-weight actors, GRSE outperformed MASkills on HotpotQA and every LoCoMo category except open-domain. From planted harmful skills, it assigned the correct sign to every planted skill and recovered a clean-library score.
Unconfirmed: Comparisons are the authors' reported results; independent replication is not stated.
- Largest reported gains were +6.7 and +3.2 F1 on multi-hop questions.
- Recovery from a corrupted library was +6.3 points versus a clean library.
- MASkills misattributed two planted skills and fell below its starting score.
- Learned skills transferred between the multi-hop tasks.
From the original article
1 min read at huggingface.co
Cooperative language-model agents can improve without weight updates by revising reusable skills, but only if a shared team reward is traced to the responsible agent and skill. Existing methods credit the skills present in each trajectory, yet presence is not responsibility: a harmful skill active in successful runs still earns positive credit.
By Yanan Ma, Yang Wang, David Le · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.