Paper turns old agent rollouts into weights, keeping 93.4% of ICL gains
burny_tech · x · 2026-07-28
Sample-Efficient Learning from Agent Experience
The paper proposes Experience Distillation for agents: instead of re-running costly environment rollouts, an experience-conditioned teacher inspects past agent trajectories and provides next-step decisions at real history points. A student is then trained to imitate those decisions without needing the experience context.
Key results reported in the abstract:
- No extra environment rollouts or long synthetic trajectories are required.
- The method retains up to 93.4% of in-context learning gains.
- It matches reinforcement-learning baselines with at least 9.6× fewer environment samples.
- On 749 curated software-engineering tasks and six text-adventure games, it preserves at least 64.8% of ICL gains across both domains.
The paper positions this as a practical way to internalize agent experience into weights while keeping sample efficiency high.
More from coding & agent
- AI agents are replacing human-centric security workflows, and Black Hat is watching — drhyrum · 2026-07-28
- GitHub project turns expert methodologies into reusable AI skills for agents — tom_doerr · 2026-07-28
- Startup building a free open-source AI harness asks what features matter most — athsrva · 2026-07-28
- ChatGPT Voice and Codex handle Amazon tasks on a locked Mac — SIGKITTEN · 2026-07-28
- Loop engineering uses a frontier model to build skills, then loops GLM 5.2 calls until it works — steipete · 2026-07-28
- Mnemos is getting a rebuilt desktop app, new MCP tools, and a redesigned platform — RileyRalmuto · 2026-07-28