MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering

MLExpert. Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering. By Shangheng Du and nine others, from East China Normal University and the Shanghai Artificial Intelligence Laboratory. Published on the fifteenth of October. Its code repository is public and real: about twenty source files.

Large language models are moving from assisting with research and engineering toward autonomously developing and improving AI systems. The authors frame this as AI for AI, a pathway toward recursive self-improvement. Machine learning engineering systems are one practical route. Workflow-driven systems manage long-horizon experimentation through complex orchestration. More recent systems let the model direct the interaction itself, organising experiments and adapting its optimisation strategy from execution feedback.

Training a model to work that way is hard for a specific reason. Feedback from any one experiment reflects several decisions that came before it, and a long run contains many such experiments, each of them costly. Training therefore has to pick useful supervision out of long interaction trajectories, and learn from intermediate outcomes rather than waiting for the end.

The paper proposes a post-training method for autonomous, long-horizon machine learning engineering, and uses it to train MLExpert, a thirty-five billion parameter Mixture-of-Experts agentic model. The central move is in how reward is assigned. Windowed Interaction RL, or WIRL, takes the multi-turn decisions made between two successive experiment results, treats them as one execution window, and gives that window a shared, outcome-based reward. Every decision inside the window is credited with what the next experiment result showed. That matches the difficulty just described: an outcome reflects several preceding decisions, so those decisions share it.

Three other pieces support it. The Kaggle-MLE Factory constructs executable tasks, and an interaction harness lets the model direct its own experimentation. Hierarchical agentic SFT transfers general tool-use and coding ability into machine learning engineering, through outcome-guided supervision. And asynchronous collection with policy correction allows the policy to be updated before a trajectory has ended, so training does not have to wait for each long run to finish.

On MLE-Bench Lite, MLExpert raises the base model's average medal rate from thirty-seven point nine percent to seventy-two point seven percent, nearly doubling it. That puts its performance in a range comparable to larger frontier models such as Kimi-K3 and GLM-5.3.

Further experiments show improvements beyond that benchmark: on NatureBench Lite, on agent scaffold optimisation, and on algorithm optimisation. The authors read this as a sign that the learned optimisation capability generalises to other tasks.

Those last results are where the word generalist in the title comes from. A skill learned on tasks built by the Factory appears to carry over to other kinds of optimisation. Underneath, the method rests on a choice of granularity. Credit is assigned neither to a single step nor to a whole run, but to the stretch of decisions between one experiment result and the next, which is the unit at which feedback actually arrives.