ForkPilot: Self-Evolving Policy for Retrospective Search in Long-Horizon Agents
Developed a self-evolving policy that learns from completed trajectories to allocate retrospective search, matching competitive performance with up to 59.2% fewer tokens on public benchmarks.