AI Papers Reader

Personalized digests of latest AI research

View on GitHub

MLExpert Learns to Run Its Own Machine Learning Experiments

Researchers at the Shanghai Artificial Intelligence Laboratory and East China Normal University report a method for training a language model to direct long machine learning engineering projects on its own. They argue the approach is a practical step toward AI systems that can help improve other AI systems. Machine learning engineering (MLE) involves building, testing and refining models against real data, and each experiment is slow and expensive. Feedback from one run reflects many earlier decisions, which makes it hard to tell a model which choices helped.

The team’s model, MLExpert, is a 35-billion-parameter “mixture-of-experts” model, meaning only a subset of its parameters (about 3 billion) is active for each token. It starts from Qwen3.6-35B-A3B. Rather than following a fixed workflow, MLExpert works inside an interaction harness, a software environment where it writes code, runs experiments, reviews results, and chooses which candidate solution to submit. Training combines two stages. The first is supervised fine-tuning on demonstration trajectories, which are recorded sequences of model actions and outcomes. The second is reinforcement learning called Windowed Interaction RL (WIRL).

WIRL groups the model’s decisions between two successive experiment results into a single “execution window” and gives every decision in that window one shared reward. The reward depends on the outcome: an execution error earns −1.0, while a gold-medal submission earns 2.0 plus a small trend term. The authors describe a concrete case from MLSP 2013 Birds, an audio classification task. MLExpert moved from audio features with LightGBM, a gradient-boosting method, through pretrained audio embeddings and fine-tuning of an audio transformer, improving ROC-AUC (a ranking score where 1.0 is perfect) from 0.58 to 0.964. The same starting point reached 0.88 with the supervised-only variant, MLExpert-SFT.

On MLE-Bench Lite, a 22-task benchmark, the share of tasks earning any medal rose from 37.9% to 72.7% under the team’s harness. Each task accounts for about 4.5 percentage points, so the gain is roughly 35 points, or nearly double the base rate in relative terms. Ablations show that general agent training raised the rate to 47.0%, supervised MLE training to 62.1%, and WIRL to 72.7%. Under the general-purpose Claude Code harness, the base model scored 42.4% and MLExpert 69.7%. The 72.7% result matches Kimi-K3 and GLM-5.3 under the team’s harness, though larger models such as GPT-5.6 Sol reached 77.3%. On NatureBench Lite, a 10-task set of scientific reproduction problems, the rate of surpassing published results rose from 0.0% to 50.0% for Claude Code runs. On a separate AIME scaffold-optimization task, a scaffold designed by MLExpert reached 43.7% held-out accuracy, compared with 23.3% for one designed by the base model.

The authors acknowledge important caveats. Results rest on 22 and 10 tasks, so single differences of a few points may not be meaningful. The fine-grained reward is coarse, since every token in a window shares one weight. Each task also allowed up to 3,000 steps and 12 hours on an H200 GPU, so the gains come with substantial compute. The team also notes that WIRL improved over supervised training but did not replace it, and that its benefits were measured on a limited set of tasks.

The work suggests that long-horizon experiment management can be trained as a model capability rather than hand-coded orchestration. Its broader significance will depend on wider and independent testing.