Sunday, 11 October 2026 · 6 min read
Recursive Self-Rewrite improves model task-solving by rewriting multi-harness trajectories
Recursive Self-Rewrite (RSR) uses a single model to collect successful task solutions under diverse execution harnesses, then rewrites these trajectories into a general harness format for supervised finetuning, boosting pass@3 rates by up to 20.8 points on complex benchmarks.
Based on the paper Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite by Zongxia Li, Yucheng Shi, Zhongzhi Li, Junyao Yang and 6 others
Harness diversity enables solving more complex tasks but complicates learning
AI models solve complex tasks through interaction with an execution harness, which controls how the model observes, acts, verifies progress, and recovers from failure. Different harnesses provide different workflow controls and verification mechanisms, enabling the same model to solve different subsets of tasks.
Improved harnesses can boost task-solving performance without changing model weights, but successful trajectories collected under specialized harnesses embed harness-specific interventions and workflow conventions. These details do not transfer directly to a general harness used at inference time, causing distribution mismatch if naively combined for training.
To improve a model's general-purpose capabilities, it is necessary to convert successful solutions discovered under diverse harnesses into reusable demonstrations compatible with a single general harness. This conversion must preserve useful planning, verification, and recovery behaviors while removing harness-specific artifacts.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short