How to improve LLM responses in domains we can’t score? Implicit signals from structured dialogue help LLM agents edit their own contexts, improving responses dramatically. “Self-evolving expertise in complex non-verifiable subject domains: dialogue as implicit meta-RL”. arxiv.org/pdf/2510.15772

New paper just out on multi-agent reinforcement learning in an open-ended environment. It introduces the RULE algorithm, allowing groups of agents to update their own reward functions to solve otherwise insoluble problems. Fixed reward functions, so 2024… www.jmlr.org/papers/volum...