ROFT: fine-tuning only on self-explanations matches GRPO on SWE-bench without RL
iScienceLuvr · x · 2026-09-29
A new arXiv paper introduces Retrospection-Only Fine-Tuning (ROFT), a minimal online procedure where an agent attempts a task, generates a retrospective explanation of its experience, and is fine-tuned with next-token prediction loss on the explanation tokens alone — no external teacher, no reward-based policy update.
Key results:
- Trained on Qwen3.5-4B using a mix of successful and failed base-model attempts
- Reaches 49.2% on SWE-bench Verified and 26.8% on SWE-bench Pro after 20 updates, no verifier needed
- Beats GRPO's 48.0% / 25.3% after 40 updates, with faster early progress in training time and sampled attempts
- Learns to solve tasks where all 64 sampled base-model attempts failed, showing learning can begin without any initially successful trajectories
More from coding & agent
- Generic API-catalog MCP server vs one tool per job: a context-bloat experiment — MandeepKaur9407 · 2026-11-16
- Berkeley paper: LLMs answer with stale values even when the update is in context — ombre-ai · 2026-11-16
- Effective HTML: 4.2k-star agent skills turn wireframes into self-contained interactive prototypes — tom_doerr · 2026-11-16
- skills-gate: one gate skill fronts 148 agent skills at just 389 tokens — Distinct-Pie2389 · 2026-11-16
- Open-weight Karpo-Decision-31B tops Decision Index at 73.9, one forward pass per answer — CodeByPoonam · 2026-11-16
- shadcn boosts @uisdk: one-shot UI rebuild that's instantly E2E-testable by AI — shadcn · 2026-11-16