PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents
PILOT lets a supervisor intervene while an agent is still working, then turns run-time failures into reusable skills.
The paper argues that long-horizon agents waste useful experience when self-improvement happens only after a run ends. Its supervisor-worker setup can redirect or abort the active worker, while also distilling procedures and failure modes into persistent memory. Across two frozen backbones and three benchmarks, PILOT ranked first in five of six configurations, including gains of up to 9.8 points on Terminal-Bench 2.0. In the self-improvement setting, it raised scores while cutting mean output tokens by more than 40%. HF Daily Papers' note
The paper argues that long-horizon agents waste useful experience when self-improvement happens only after a run ends. Its supervisor-worker setup can redirect or abort the active worker, while also distilling procedures and failure modes into persistent memory. Across two frozen backbones and three benchmarks, PILOT ranked first in five of six configurations, including gains of up to 9.8 points on Terminal-Bench 2.0. In the self-improvement setting, it raised scores while cutting mean output tokens by more than 40%. HF Daily Papers' note
score 5