Researchers have introduced Recursive Self-Rewrite (RSR), a framework designed to convert successful solutions found under specialized task harnesses into reusable training trajectories for a model operating under a general harness. The work addresses a practical gap: harnesses that help models solve difficult tasks often introduce interventions that may not be available at deployment time.
RSR relies on a single base model, Qwen-3.8-27B, and divides the work among three roles. A planner extracts procedures into runbooks, a critic screens for verifier and solution leakage while guiding recursive revision, and an executor follows qualified runbooks in fresh sandboxes.
Across roughly 3,000 self-curated terminal tasks, three harnesses jointly solved 759 tasks, which the authors report as 34.3% more than the strongest individual harness in the recorded pool. From 2,001 successful source trajectories, RSR expanded the data into 11,094 rewritten trajectories for supervised finetuning.
Training on these rewritten trajectories outperformed both the base model and direct trajectory SFT. Compared with the base model, pass@3 rose from 57.0% to 74.2% on Terminal-Bench 2, from 1.5% to 9.1% on Terminal-Bench 4, from 39.0% to 63.0% on the self-curated Terminal-Bench Hard, and from 3.0% to 6.0% on the self-curated Software Terminal-Bench. Process reward on Long-Horizon Terminal-Bench increased from 0.21 to 0.29.
The authors frame the results as evidence that diverse, harness-assisted experiences can be reconstructed into capabilities that a model can reuse under a general harness.