@dair_ai
Who should design the training environment for an RL agent, the practitioner or the policy itself? RL pipelines for LLMs usually rely on manually redesigned environments between stages, with practitioners guessing which configuration will best improve the current policy. This work proposes an LLM-as-Environment-Engineer framework. The current policy analyzes its own failure trajectories plus context and proposes the next-stage environment configuration, automating a step that has stayed stubbornly manual. They also release MAPF-FrozenLake, a controllable multi-agent testbed whose generator exposes multi-dimensional environment configs. Why does it matter? Curriculum design between RL stages is mostly gut feel today. Letting the policy read its failures and shape the next environment closes a loop that practitioners currently close by hand. Paper: https://t.co/lZHlqozrQD Learn to build effective AI agents in our academy: https://t.co/LRnpZN7L4c