Researchers alternate agent sessions on machine-learning tasks
Humanize MLE Agents alternates two agents, resetting session context while preserving a shared workspace.
Event / data period: 2026-10-09
Published: · Updated:
Researchers affiliated with Tsinghua, Zhejiang and NVIDIA presented Humanize MLE Agents, a method that alternates two agents’ sessions on long machine-learning engineering tasks. Each session stops after a capped number of experiments; context resets between sessions while code, models and prior results remain in a shared workspace.
Across 75 Kaggle tasks in MLE-bench, the authors report a 78.2% any-medal rate and say the workflow surpassed the strongest open-source comparison with a smaller time budget. This is a result for a specific competition set and agent configurations; it does not guarantee the same gain across model versions or real engineering projects.
Data
Measure
75 Kaggle tasks; MLE-bench evaluation
| Any-medal rate on MLE-bench Kaggle tasks | 78.2 | 75 Kaggle tasks; MLE-bench evaluation |
Methodology
Values come from the paper's stated benchmark experiments; they are not field measurements or general performance guarantees.
Updated:
Measure
Kaggle task set reported in the MLE-bench evaluation
| Number of MLE-bench Kaggle tasks | 75 | Kaggle task set reported in the MLE-bench evaluation |
Methodology
Values come from the paper's stated benchmark experiments; they are not field measurements or general performance guarantees.
Updated:
Subjects and places
Subject proposals
- Science and technologySuggested
- Preserved editorial category/topic; broad candidate onlyarxiv.org ↗
Geographic references
Only source-provided geographic references are shown. Opening an external map is optional. Optional story maps use OpenStreetMap. Opening a map sends your IP address and this site’s address to the map provider; location data remains readable as text without loading the map.
- China · Place role: Coverage · Place type: Country or area · Location precision: RegionalEvidence
- Explicit preserved editorial country/region; no city or national centroid inferredarxiv.org ↗