AI
MILO: Automated Harness Discovery via Orchestrated Multi-Agent Evolution
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 02/10/2026, 08:44:40 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 01/10/2026, 17:18:24 UTC.
Briefings
MILO is a framework for automated harness discovery using hierarchical lineage memory and per-island mutator agents. It outperforms existing methods on Terminal-Bench 2.1, PaperBench, and DeepSWE, achieving 86.1 pm 2.0% on Terminal-Bench 2.1 compared to the official leaderboard's top entry of 83.8 pm 2.3%. MILO also improved Erdős minimum-overlap and autocorrelation inequalities on EinsteinArena open problems. The study utilized Opus 4.8 and gpt-oss-120b models across multiple benchmarks.
Key points
- MILO uses hierarchical lineage memory over island-based trees, where rejected mutations serve as negative evidence for the search process.
- MILO achieves 86.1 pm 2.0% on Terminal-Bench 2.1, exceeding the official leaderboard's top entry of 83.8 pm 2.3% while using 26% fewer tokens than its initial harness.
Context and limitations
Results are limited to three benchmarks (Terminal-Bench 2.1, PaperBench, DeepSWE) and two open problems on EinsteinArena. Performance gains are measured against initial harnesses and prior search methods, not necessarily against all possible future harnesses. Specific agent behaviors and mutation rates are not quantified in the provided text.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 02/10/2026, 08:44:40 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeic4xjlb2bbgse33bjtjvp6nk2gepctuzyfb4dqcjb5kfuxf6n3i7m