Meta's RSD paper: agents self-improve from unlabeled runs, lifting Qwen3-14B success from 12.9% to 20.6%
rohanpaul_ai · x · 2026-10-23
Meta researchers propose Retrospective Self-Distillation (RSD), a multi-round framework that lets agents learn from their own past task attempts even when nobody marked them right or wrong.
- The same model acts as both student and teacher: the teacher re-scores every action in a finished run, conditioned on the outcome, cross-task precedents, and hindsight
- Refresh the retrospective context each round — frozen notes left ALFWorld success at 59.7% by round 3 vs 67.9% with fresh reviews
- Results: Qwen3-14B rose from 12.9% to 20.6% on ScienceWorld; Qwen3-8B hit 76.6% on ALFWorld, beating methods given rewards or ground truth; up to 59.7% relative gain over base, plus 44.7% more with frontier-guided context-policy evolution
Takeaway: keep your agent's unlabeled runs, failures included, and train on hindsight reviews of them.
More from coding & agent
- Most new CVEs hit hand-written legacy code, dev pushes back on vibe coding security blame — HowDevelop · 2026-10-23
- Claude and Claude Code Merge: Anthropic Engineers Say Delete Your CLAUDE.md Every 6 Months — rubenhassid · 2026-10-23
- Engineers merging 4x more PRs at Imprint: Will Larson on speeding up code review — AxSaucedo · 2026-10-23
- 15 practical ways to hand off work to Grok Bot's cloud computer — even with laptop closed — XFreeze · 2026-10-23
- 3 Lessons from Rebuilding Compute for AI Agents: Safe Sandboxes, Near-Free Idle, Persistent Memory — Saboo_Shubham_ · 2026-10-23
- Local model agentic coding benchmark: 27B matches 125B, tool-calling is the lever — Joralk_ · 2026-10-23