Can language agents learn lessons that transfer to new tasks?

Most language agents can only fix mistakes within a single task. This work asks whether they can extract general lessons that transfer to entirely new problems. Researchers built a system where a reflector model watches an actor agent fail, then writes improved prompts that help the actor succeed on future tasks. Using reinforcement learning, they trained reflectors from scratch on ALFWorld and MiniHack environments. The trained reflectors beat baselines on held-out tasks and sometimes generalize to radically different environments. They also released MetaGym, a library for building and testing self-improving agents.