Research2026-09-29
Researchers proposed a method for recursive self-improvement using multiple evolving branches. Each branch keeps its own development cases and proposal policy. A router selects the best branch for new inputs. The system outperformed Meta-Harness on math reasoning by 34.8%. It also beat Meta-Harness on Terminal-Bench 2.0 by 11.6%. SWE-bench Lite scores rose by 3.8%. Selection and routing used only development data. The paper was submitted to arXiv on 29 September 2026.
What changed
Meta-Harness used a fixed development set and a single proposal policy for recursive self-improvement.
What it unlocks
Creating multiple evolving branches of agent harnesses that specialize and combine via a router.
- 34.8% improvement on Olympiad-level math
- 11.6% improvement on Terminal-Bench 2.0
- 3.8% improvement on SWE-bench Lite
Sources