Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite

Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite. By Zongxia Li and nine others, from Tencent HY LLM Frontier, the University of Maryland, College Park, and five other institutions. Published on the second of October. No code repository is linked to it.

A successful trajectory on a difficult task is valuable supervision. The catch is where such trajectories tend to come from. Specialised harnesses help a model succeed, but they do so through interventions that may not be available once the model is deployed. The supervision, in other words, is gathered under conditions the deployed model will not share.

Recursive Self-Rewrite, or RSR, keeps the successes and rewrites the circumstances. One base model, Qwen-3.8-27B, first discovers successful solutions under a variety of harnesses. The same model then reconstructs those solutions as training trajectories under a general harness, the kind it will actually operate in.

Three roles do the reconstruction. A planner extracts the procedure behind a solution into a runbook. A critic screens each runbook for leakage, of the verifier or of the solution itself, and guides recursive revision until the runbook qualifies. An executor then follows qualified runbooks in fresh sandboxes, producing new trajectories that never relied on the original harness.

The task pool holds roughly three thousand terminal tasks, curated by the model itself. Three harnesses together solve seven hundred and fifty-nine of them, thirty-four point three percent more than the strongest single harness in the pool. RSR expands two thousand and one successful source trajectories into eleven thousand and ninety-four rewritten ones for supervised fine-tuning. Training on them beats both the base model and direct fine-tuning on the original trajectories.

Measured by pass at three against the base model, Terminal-Bench 2 rises from fifty-seven percent to seventy-four point two percent. Terminal-Bench 4 rises from one point five percent to nine point one. The authors' own Terminal-Bench Hard goes from thirty-nine percent to sixty-three, and their Software Terminal-Bench doubles, from three percent to six. Process reward on Long-Horizon Terminal-Bench climbs from zero point two one to zero point two nine.

The result reframes what a harness is for. Rather than a crutch the model has to carry into deployment, it becomes a way of finding solutions that are then rewritten into a form the model can follow unassisted. The authors frame it as turning diverse harness-assisted experience into reusable capability for a model running under a general harness, and the comparison with direct training on the original trajectories is what shows the rewriting step earning its place.