STAIR hits 81.2% on SWE-bench by reusing old fixes

STAIR turns old repair trajectories into reusable, multi-level plans and hits 81.2% Pass@1 on SWE-bench Verified with MiniMax M2.5, per the paper posted to arXiv.

That's ahead of the 79.2% GPT-5 gets on the same benchmark. Most repair agents treat every issue independently, discarding whatever procedural knowledge they picked up solving the last one.

The system builds a tree instead of starting fresh. Fine-grained diagnostic actions sit at the bottom, high-level strategies at the top, and it draws from both to write issue-specific prompts.

The transfer result matters most. Researchers took the same plans, made zero code changes, and dropped them into mini-SWE-agent v2, a structurally different agent. Its score jumped from 75.8% to 81.0%.

It's not automatic. Ablations show mixing abstraction levels beats sticking to any single level, and raw, unabstracted trajectories transfer far worse than the distilled plans do.

For builders running repair agents, the lesson is to keep old trajectories instead of discarding them. Abstract them into levels rather than feeding raw logs back in.