Humanize MLE Agents: Multi-Agent Session Alternation Beyond Single-Agent Limits
Humanize MLE Agents alternates two agents and hits 78.2% any-medal on 75 MLE-bench Kaggle tasks.
The paper models alternating agent sessions as a context-reset semi-Markov decision process and introduces Humanize MLE Agents (HMA), which swaps two agents on a shared workspace, caps each session at k experiments, and resets context without task-specific harnesses. On 75 Kaggle tasks in MLE-bench, HMA reports a 78.2% any-medal rate, 8.0 points above the strongest open-source baseline with a quarter of its time budget. Across six configurations spanning five heterogeneous agent pairs, the average gain over constituent baselines is 6.4 points. A kernel-optimization case reaches 878 cycles, within 1.6% of the community leaderboard leader.
- HMA alternates two agents on a shared workspace with per-session experiment caps and context resets.
- It reaches 78.2% any-medal rate on 75 MLE-bench Kaggle tasks, 8.0 points above the strongest open baseline.
- That result used about a quarter of the baseline time budget.
- Across six configurations and five agent pairs, average any-medal gain was 6.4 points.
- A kernel-optimization case hit 878 cycles, within 1.6% of the community leader.
From the original article
1 min read at arxiv.orgExisting LLM-based agent approaches for long-horizon machine learning engineering (MLE) often use context engineering, implemented through custom harnesses that manage tools, skills, and accumulated experience.
By Changye Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.