Which Agent's Skill Should Change? Group-Relative Credit Assignment for Evolving Multi-Agent Teams

Which Agent's Skill Should Change? Group-Relative Credit Assignment for Evolving Multi-Agent Teams. By Yanan Ma and eight others, from the Mayo Clinic, the University of Manchester, and two other institutions. Published on the nineteenth of October. Its code repository is worked out: about one hundred and ten source files, with a readme and a licence.

A team of cooperating language-model agents can get better without any weight updates, by revising the reusable skills each agent draws on. That only works if a reward shared by the whole team can be traced back to the agent and the skill responsible for it. Existing methods credit whichever skills were present in a trajectory. Presence, though, is not responsibility. A harmful skill that happens to be active during successful runs still collects positive credit, and the team keeps it.

GRSE, short for Group-Relative Skill Evolution, assigns credit by comparison instead of by presence. The inspiration is GRPO. Rather than judging one run on its own, the method looks at a group of runs of the same question and asks which differences go with better and worse outcomes.

Concretely, GRSE executes each question several times under the same skill libraries, then contrasts those executions to credit individual agent and skill pairs. Failures are not just scored. They receive a structured diagnosis. The credit that results is behavioural: it reflects what a skill did in the runs being compared, not merely the fact that it was there.

Evaluation uses HotpotQA and LoCoMo with two open-weight actors. GRSE beats MASkills, the closest prior method, on HotpotQA and on every LoCoMo category except open-domain. The largest gains come on multi-hop questions, at six point seven and three point two F1. Those gains rest on better-grounded retrieval rather than on answers given without evidence, and the learned skills transfer between the multi-hop tasks.

The sharpest test plants harmful skills in the starting library. GRSE credits every planted skill with the correct sign and recovers the score of a clean library, a gain of six point three points. MASkills misattributes two of the planted skills and ends below where it started.

The corrupted-library experiment is the one that matches the title's question. When bad skills sit alongside good ones in winning runs, crediting by presence cannot tell them apart. Comparing repeated runs of the same question can, and that difference decides whether a self-revising team cleans up its library or entrenches its mistakes.