RHO improves LLM agents from unlabeled past trajectories

Summary

RHO (Retrospective Harness Optimization) is a method that improves LLM agents by learning from unlabeled past trajectories. The approach leverages previous execution records to enhance agent performance.