HMA
A dual-agent alternating collaboration framework that enables multiple agents to complete long-horizon machine learning engineering tasks in relay, using a shared workspace and inter-session context reset mechanisms, without requiring task-specific harnesses or prior knowledge injection.
Timeline (last 90 days)
HMA (Humanize MLE Agents) has two agents work alternately on the same shared workspace, executing at most k experiments per session, resetting context between sessions, and relying on no additional task-specific harnesses or injected prior knowledge.
HMA achieved an any-medal rate of 78.2% on the MLE-bench benchmark (covering 75 Kaggle tasks), setting a new state-of-the-art (SOTA).
Compared to the strongest open-source baseline method, HMA improved the any-medal rate by 8.0 percentage points while using only one-quarter of the time budget.
Across six configurations covering five groups of heterogeneous agent combinations, HMA improved the any-medal rate by an average of 6.4 percentage points over its constituent baseline agents.
When HMA is applied to a kernel optimization scenario, HMA reaches 878 optimization cycles, with a gap of only 1.6% from the top of the community leaderboard.
The research team established sufficient conditions for an alternating policy to surpass the single-agent reference upper bound based on CR-SMDP, and derived the experimental budget boundary.
All Facts (6)
HMA (Humanize MLE Agents) has two agents work alternately on the same shared workspace, executing at most k experiments per session, resetting context between sessions, and relying on no additional task-specific harnesses or injected prior knowledge.
50%UnverifiedHMA achieved an any-medal rate of 78.2% on the MLE-bench benchmark (covering 75 Kaggle tasks), setting a new state-of-the-art (SOTA).
50%UnverifiedCompared to the strongest open-source baseline method, HMA improved the any-medal rate by 8.0 percentage points while using only one-quarter of the time budget.
50%UnverifiedAcross six configurations covering five groups of heterogeneous agent combinations, HMA improved the any-medal rate by an average of 6.4 percentage points over its constituent baseline agents.
50%UnverifiedWhen HMA is applied to a kernel optimization scenario, HMA reaches 878 optimization cycles, with a gap of only 1.6% from the top of the community leaderboard.
50%UnverifiedThe research team established sufficient conditions for an alternating policy to surpass the single-agent reference upper bound based on CR-SMDP, and derived the experimental budget boundary.
50%