AI
Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 01/10/2026, 20:00:00 UTC
Briefing created: 05/10/2026, 02:38:27 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 02:18:24 UTC.
Briefings
Recursive Self-Rewrite (RSR) uses a single base model to discover successful solutions across diverse harnesses and reconstruct them as training trajectories. This approach expands 2,001 successful source trajectories into 11,094 rewritten trajectories, outperforming individual harnesses and direct SFT. Key metrics include a 74.2% pass@3 increase on Terminal-Bench 2 and a 0.29 process reward rise on Long-Horizon Terminal-Bench compared to the base model.
Key points
- RSR expands 2,001 successful source trajectories into 11,094 rewritten trajectories for supervised finetuning.
- RSR achieves a 74.2% pass@3 increase on Terminal-Bench 2 compared to the base model.
Context and limitations
Results are based on approximately 3K self-curated terminal tasks across three harnesses; specific deployment constraints may limit generalizability.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 05/10/2026, 02:38:27 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibukytclf4du2a7zpgsa65ljcujq6d234h5jvzqiqyir6xtvum7ty