The authors study teams of two open-weight language-model agents, one retrieving evidence and one reasoning, where each agent's behavior is controlled by natural-language skills and model weights stay fixed.
Contrasting group executions pinpoints which agent skill caused success or failure
GRSE compares multiple rollouts of the same question to credit agent-skill pairs, beating prior MASkills on multi-hop QA and pruning harmful skills that MASkills retains.
Chinese Tech
Yanan Ma · Yang Wang · David Le · Chenhan Yuan · Yizheng Sun · Guangwei Zhang · +3 more
Mayo Clinic · University of Manchester · Alibaba Token Hub · Alibaba Group · Shanghai University of Finance and Economics
Research Digest··3 min read
The authors introduce GRSE, a training-free framework for evolving cooperative language-model agents.
§