Back to AI Research

AI Research

MLExpert learns from experiment windows in autonomous machine-learning work

Key Takeaways

  • MLExpert combines supervised training and window-based reinforcement learning, reporting a higher medal rate on MLE-Bench Lite and tests of transfer to other tasks.
  • Machine-learning engineering involves a sequence of costly experiments, each reflecting several earlier choices.
  • The [MLExpert paper](https://arxiv.org/abs/2610.22344) studies how to train an agent from those intermediate outcomes, rather than waiting for one result at the end of a long optimization run.
  • ## Train for model-directed experimentation
  • MLExpert is a 35B mixture-of-experts agentic model.

Machine-learning engineering involves a sequence of costly experiments, each reflecting several earlier choices. The MLExpert paper studies how to train an agent from those intermediate outcomes, rather than waiting for one result at the end of a long optimization run.

Train for model-directed experimentation

MLExpert is a 35B mixture-of-experts agentic model. The researchers build a Kaggle-MLE Factory to create executable tasks and a harness in which the model organizes experiments and adapts its strategy from execution feedback.
Their supervised stage transfers general coding and tool-use capabilities into machine-learning engineering. The subsequent Windowed Interaction RL method groups the decisions between successive experiment results into an execution window.
Those decisions receive a shared outcome-based reward. Asynchronous collection and policy correction allow updates before a whole trajectory finishes.

Read the reported benchmark improvement

The authors report an average medal rate of 72.7% on MLE-Bench Lite. The corresponding base-model result is 37.9%, and the paper compares the trained model with larger frontier systems.
Medal rate belongs to the benchmark's evaluation procedure. It does not mean the system produces a deployment-ready model for that share of arbitrary business problems, nor does it replace checks for leakage, reproducibility or resource use in another setting.
The authors also report improvements on NatureBench Lite, agent scaffold optimization and algorithm optimization.

Evaluate the experiment process

The training method uses windows between experiment results. This is the paper's practical contribution for researchers designing long agent trajectories: intermediate feedback may offer useful supervision while avoiding a wait for every experiment in a run.
A follow-up evaluation should inspect which experiments the agent selects, how it accounts for unsuccessful runs and whether the gains persist under a fixed budget. The paper frames these findings within autonomous MLE research; they should not be expanded into a demonstrated ability for open-ended recursive self-improvement.

Comments