Skip to content

Back to the catalog

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

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

Presents MLExpert, a 35B Mixture-of-Experts agentic model trained for autonomous long-horizon machine learning engineering. Training uses a Kaggle-MLE Factory to build executable tasks and an interaction harness, hierarchical agentic SFT with outcome-guided supervision, and Windowed Interaction RL (WIRL). WIRL treats the multi-turn decisions between successive experiment results as one execution window with a shared outcome reward, and uses asynchronous collection and policy correction. On MLE-Bench Lite the average medal rate rises from 37.9% to 72.7%, comparable to larger frontier models such as Kimi-K3 and GLM-5.3. It also improves on NatureBench Lite, agent scaffold optimization and algorithm optimization.

TL;DR MLExpert, a 35B MoE model post-trained with agentic SFT and windowed-interaction RL, raises MLE-Bench Lite medal rate from 37.9% to 72.7%.
Open original ↗
Document

What it claims

MLExpert is a 35B Mixture-of-Experts model post-trained so that the model itself drives a long machine learning engineering run. It decides when to analyse, revise, execute, backtrack or switch direction, with no external workflow or search controller. The authors contrast this with systems such as ML-Agent (multi-turn demos plus single-step RL) and Frontis-MA1 (operators trained to plug into an external evolutionary search). Those train single-step or local operations, while MLExpert trains the whole optimization policy. On MLE-Bench Lite the base model's average medal rate rises from 37.9% to 72.7%, which the authors call comparable to Kimi-K3 and GLM-5.3.

How it works

Kaggle-MLE Factory. A multi-agent pipeline converts historical Kaggle competitions into executable tasks:

  • A discovery agent filters competitions by status, scale and submission format.
  • An analyzer agent extracts the task spec and modality.
  • A generator agent builds a hidden-label test split from the original labelled data and writes the evaluator.
  • A validator repairs data-integrity, format and scoring problems.
  • A final manual review drops tasks with unresolved label or scoring issues.

Tasks span table, image, text and audio data.

Interaction harness. This is a ReAct-style loop. State lives in an external workspace as solution nodes (code, logs, metadata), and parent-linked revisions form an experiment tree that the model can revisit. Named tools:

  • execute_code runs an experiment.
  • update_answer sets the submission candidate.
  • analyze delegates to a sub-agent with its own context and returns only a summary.
  • invalidate_node excludes candidates the model suspects of overfitting or leakage.

When the context fills up, older messages are summarized and recent ones are kept.

Hierarchical SFT, two stages.

  1. General coding and tool-use trajectories.
  2. MLE teacher trajectories plus literature-derived scientific tasks.

Stage 2 uses initial-solution prefixes, full optimization trajectories and analysis sub-agent traces. Outcome-guided selection works like this:

  • Find the first best submission.
  • If performance degrades past a threshold after it, truncate the trajectory at that peak. Otherwise keep it through the last submission.
  • Compute loss only on model reasoning, text and tool-call tokens.

WIRL. Each execute_code result closes an "execution window" made up of all the decisions since the previous result. Every decision in that window gets a shared reward. Medal thresholds normalise across metrics:

OutcomeReward
Execution error caused by the model−1.0
No error, but no scorable submission−0.6
Valid, below medal−0.4 + T
Bronze or silver1.0 + T
Gold2.0 + T

T is a trend term compared with the previous valid submission: −0.2 if worse, 0 if unchanged, +0.2 if better, and 0 for the first valid submission. Windows are collected asynchronously, and policy correction lets updates happen before a trajectory finishes. The loss is window-normalized.

Evidence

The fetched text gives only the headline MLE-Bench Lite figure: average medal rate of 37.9% for the base model versus 72.7% for MLExpert. It also makes qualitative claims of gains on NatureBench Lite, agent-scaffold optimization and algorithm optimization. Evaluation follows the MLE-bench protocol: the agent sees the task description and public data, builds its own validation, and the final submission is scored on hidden labels within a time budget.

How to use it

The page links to code and model artifacts. The text does not name them, and it gives no install commands, weight names, licence, or compute and time budget.

Limits

  • The fetched text stops partway through the WIRL section. Ablations, per-task tables, baseline settings, compute, the time budget B and the base model's identity are not visible here.
  • "Comparable to Kimi-K3 and GLM-5.3" comes without numbers in this text.
  • Generalization beyond MLE is asserted, but no scores are shown.
  • Several thresholds are not given in the visible text: the SFT quality threshold, the degradation threshold and the medal cutoffs. They are deferred to Appendix C.1.
  • Factory tasks come from historical Kaggle competitions, as MLE-Bench tasks do. The visible text does not say how overlap between training tasks and MLE-Bench Lite is handled. Check this before trusting the jump in medal rate.

Organisations named here

  • arXiv
  • MLE-Bench
  • NatureBench

Technologies named here

  • Kaggle-MLE Factory
  • MLExpert
  • Kimi K3
  • GLM-5.3
  • LLM
  • Mixture of Experts

Techniques named here

  • Hierarchical agentic SFT
  • Windowed Interaction RL
  • Asynchronous collection and policy correction
  • Model-directed interaction

Related in research