Mixture of Self-Improving Branches for Agent Harness Optimization¶
L1 mechanism diagnostic. Branch subset differentiation and input routing, no LLM-generated harness or SWE-bench run.
Paper information¶
| Field | Value |
|---|---|
| Paper | arXiv 2609.37834 |
| Affiliations | No first-author affiliation listed on the paper's first page |
| First public release | 2026-09-29 (arXiv v1) |
| Authors' code | No public author repository found as of 2026-10-01. |
| Adapter / method | branch-mixture |
| Local code | src/auto_research/agent_research/latest_20261001.py |
Original paper¶
Background and contribution¶
Maintain separate development subsets and proposal policies. Remove examples every branch solves and retain examples that distinguish branches. At deployment an input-only router selects a development-set champion.
flowchart LR
I[Current state] --> B[Branch proposals and subsets] --> D[Keep discriminative examples] --> R[Input-only router]Key figure from the paper¶
Original Figure 1: branching harness optimization. From the paper; copyright belongs to the authors. Not a local result.
Core mechanism¶
\(J_b^t(H)=|X_b^t|^{-1}\sum_x\mathbb E[r(f_M(H,x),x)]\). The local update retains discriminative examples; the linear router cannot read test labels.
Paper results¶
Reported relative gains over Meta-Harness: 34.8% olympiad reasoning, 11.6% Terminal-Bench 2.0, 3.8% SWE-bench Lite.
Local reproduction¶
Compare disabled/default against enabled. Three-seed diagnostics are diagnostic_only=true, not cross-model effects.
Reproduction limits¶
Branch subset diversification and routing, without LLM harness generation or SWE-bench.