Which Agent's Skills Should Change? Group-Relative Credit Assignment in Evolving Multi-Agent Teams

GRSE (Group-Relative Skill Evolution) repeatedly executes the same question multiple times under the same skill library, and by comparing different execution results, assigns behavioral credit to agent–skill combinations rather than only based on whether a skill appears in the trajectory. The authors state that on two open-weight agents, the method outperforms the prior MASkills on HotpotQA and on all LoCoMo categories except open-domain, with F1 on multi-hop questions improving by 6.7 and 3.2 respectively; the authors state that the gains come from more solid retrieval rather than answering without evidence, and that the learned skills also transfer across multi-hop tasks. In experiments where harmful skills were implanted into the skill library, GRSE assigned the correct credit sign to each implanted skill and restored the score to the level of the clean skill library (+6.3 points), whereas MASkills misjudged two of them and its score fell below the starting point.