Search papers, labs, and topics across Lattice.
This paper introduces Recursive Self-Rewrite (RSR), a framework that leverages a single base LLM to discover and reconstruct successful solutions from specialized harnesses into generalizable training trajectories. RSR addresses the challenge of deploying specialized solutions by using a planner, critic, and executor to extract, filter, and revise procedures into runbooks executable in general environments. The approach significantly increases task completion rates and improves downstream performance on complex benchmarks compared to direct supervised fine-tuning.
LLMs can recursively rewrite their own solutions from specialized environments into generalizable training data, boosting performance on complex tasks by over 34% and improving pass@3 by up to 17 percentage points.
Successful trajectories on difficult tasks provide valuable supervision for model improvement, but specialized harnesses introduce interventions that may be unavailable during deployment. We propose Recursive Self-Rewrite (RSR), a framework that uses one base model, Qwen-3.8-27B, to discover successful solutions under diverse harnesses and reconstruct them as training trajectories under a general harness. A planner extracts procedures into runbooks, a critic screens for verifier and solution leakage and guides recursive revision, and an executor follows qualified runbooks in fresh sandboxes. Across approximately 3K self-curated terminal tasks, three harnesses jointly solve 759 tasks, 34.3% more than the strongest individual harness in the recorded pool. RSR expands 2,001 successful source trajectories into 11,094 rewritten trajectories for supervised finetuning. Training on these trajectories outperforms both the base model and direct trajectory SFT. Compared with the base model, pass@3 increases 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 our self-curated Terminal-Bench Hard, and from 3.0% to 6.0% on our Software Terminal-Bench. Process reward on Long-Horizon Terminal-Bench rises from 0.21 to 0.29. These results show how diverse harness-assisted experiences can be reconstructed into reusable capabilities for a model operating under a general harness.
