Self-Evolving Agents through Privileged Retrospective Distillation
- Published
- Source
- arXiv
- Paper number
- 1357
- Field
- AI Agents
- arXiv ID
- 2610.27495
Key points
- RSD, without ground-truth trajectories or verifier rewards, has the same model re-score its own actions under privileged retrospective context and distills the resulting distribution back into the policy.
- The privileged context combines four components: precedents and lessons from accumulated experience, plus reflection and future hindsight available only after the current trajectory ends.
- With Qwen3-8B and 14B across three benchmarks, RSD improved performance by up to 59.7% relative over the base model, and frontier-guided context-policy evolution added up to a further 44.7% relative improvement over a fixed context policy.
- On ALFWorld-8B, RSD achieved an average SR of 76.6 against 62.9 for Base, and on ScienceWorld it improved from 15.6 to 23.4 at 8B and from 12.9 to 20.6 at 14B.
- Frontier-guided context-policy evolution improved ScienceWorld SR from 23.4 to 27.5 at 8B and from 20.6 to 29.8 at 14B, and outperformed random search at both scales.
- As a boundary case, SearchQA stayed within a narrow performance band across rounds.
Paper links
External research summaries. These are not HDATF publications or measured product results.