SEAL: Synergistic Co-Evolution of Agents and Learning Environments

Researchers introduced SEAL, a closed-loop training framework that co-evolves tool-use agents and their learning environments.

It analyzes failed rollouts, turns them into labels, and uses them to update both the agent policy and the supervision environment.

Across multiple backbones, SEAL delivered substantial accuracy gains, strong out-of-distribution transfer, and better generalization.

The approach tackles agent-environment misalignment directly and achieves large improvements with only about 400 samples.