On Terminal-Bench 2, Qwen-3.8-27B sits at 57.0% pass@3. Swap in a stronger execution harness — the scaffolding that manages observations, tool calls, verification, retries and stopping — and the model suddenly solves more tasks. The catch: those extra gains are booked to the harness, not the model. Remove the scaffolding at deployment and the score goes back down.
Borrowed Experience Needs Translation First
That translation is exactly what Recursive Self-Rewrite (RSR), from Tencent HY LLM Frontier, the University of Maryland and other institutions, does. The team ran the same Qwen-3.8-27B under three harnesses (Terminus 2, Recursive Self-Reflect Terminus, StateM) across roughly 3,000 self-curated terminal tasks. Together the three solved 759 tasks, 34.3% more than the strongest single harness. But trajectories recorded under specialized harnesses are laced with controller interventions and workflow conventions; training on them naively teaches the model to lean on external crutches.
RSR splits the model itself into three roles. A planner distills procedures from successful trajectories into runbooks. A critic screens for verifier and solution leakage — leaking candidates are rejected and recursively regenerated with feedback. An executor replays each qualified runbook in a fresh sandbox under a general harness. The result: 2,001 source trajectories expanded into 11,094 rewritten, crutch-free ones.
The Scorecard, and the Cold Water
Fine-tuning on those trajectories lifts Terminal-Bench 2 pass@3 from 57.0% to 74.2%, Terminal-Bench 4 from 1.5% to 9.1%, the self-curated Terminal-Bench Hard from 39.0% to 63.0%, Software Terminal-Bench from 3.0% to 6.0%, and the Long-Horizon process reward from 0.21 to 0.29 — beating both the base model and direct trajectory SFT.
Two buckets of cold water. First, on the hardest Terminal-Bench 4 the number is still single-digit, and the long-horizon process reward only crawls from 0.21 to 0.29. Second, the paper's experiments show that direct SFT without rewriting improves some benchmarks while degrading others through looping behaviors — which is precisely the counter-evidence that rewriting, not data volume, is the load-bearing step. Also note: RSR-27B weights are on Hugging Face under Apache-2.0, but the model card README is still empty — anyone aiming to reproduce will be reading the 16-page paper themselves.
Not the Same Thing as Hyra
Readers who remember Hunyuan's July release of Hyra-1.0, the recursively self-improving research agent, may find the naming familiar. The difference: Hyra is an agent product that improves itself recursively at runtime; RSR is a training-time data-engineering method that spends the improvement on SFT data, letting experience settle into weights. One evolves while running; the other amortizes experience into the model.
As evaluation harnesses keep getting stronger, leaderboard scores increasingly read like a joint signature of model plus scaffolding. RSR offers a clean path: borrowing experience is fine, but it only counts as yours once it is rewritten to travel without crutches. Paper: arXiv:2610.02826; weights: IntelligenceLab/RSR-27B.