SEAL: Synergistic Co-Evolution of Agents and Learning Environments

In three linesSEAL is a closed-loop co-evolution framework for tool-use LLM agents. It collects verifiable trajectories, diagnoses turn-level failures, and uses these signals to jointly adapt the learning environment and agent policy. With 400 training samples, SEAL achieves +8.25 to +26.25 point gains across three backbones and shows positive out-of-distribution transfer.

Summary generated by Claude — human-verified