MLExpert AI Nearly Doubles Score to 72.7%, Rivals Kimi-K3

Autonomous machine learning engineering took a measurable step forward this week with a new paper detailing MLExpert, a 35-billion-parameter agentic model that lifted its medal rate on a standard engineering benchmark from 37.9% to 72.7%.

Key Takeaways

MLExpert raised its MLE-Bench Lite medal rate from 37.9% to 72.7% after training
MLExpert is a 35-billion-parameter Mixture-of-Experts model that routes tasks through only part of its parameters
Windowed Interaction RL assigns a shared reward to decisions made between two experiment results
The paper reports gains on NatureBench Lite and agent scaffold optimization

The researchers said in the paper, published Oct. 15, that MLExpert is a Mixture-of-Experts model, meaning it routes each task through only a portion of its total parameters rather than activating the whole network every time.

On MLE-Bench Lite, a benchmark for autonomous machine learning engineering that scores how well an AI system can plan, run and refine machine learning experiments on its own, MLExpert’s performance after training nearly doubled compared with its untrained base model.

It reached results the paper says are comparable to larger frontier models including Kimi-K3 and GLM-5.3. The paper frames this as a step toward what its authors call AI4AI, or AI systems built to design and improve other AI systems, a pathway researchers describe as recursive self-improvement.

Long-horizon machine learning engineering runs involve dozens of costly experiments, where feedback from any single result depends on several earlier decisions stacked on top of each other.

The central challenge the researchers tackled was teaching a model which of those earlier decisions actually mattered, rather than treating every step in a long run as equally important.

Also Read: Why Your AI Agent’s Benchmark Score Hides A Real Failure Rate

Why Autonomous Machine Learning Engineering Rewards Whole Experiment Windows

Reinforcement learning is a training method where a model gets a reward signal for actions that lead to good outcomes, then adjusts its behavior to repeat them. Standard reinforcement learning struggles on long tasks because a single reward at the end does not tell the model which of 50 earlier decisions actually helped.

MLExpert’s researchers built what they call Windowed Interaction RL, which groups the back-and-forth decisions between two experiment results into a single block and assigns that whole block a shared reward.

This lets the model learn from intermediate outcomes across a long run instead of waiting for one final score, the specific mechanism behind the jump from 37.9% to 72.7% on MLE-Bench Lite.

From Workflow Scripts To Self-Organizing Agents

Machine learning engineering automation started with rigid, workflow-driven systems that followed a fixed script of steps a human had designed in advance. Newer systems, including MLExpert, let the model organize its own experiments and change strategy based on what earlier runs reveal, a shift the paper calls model-directed interaction.

The researchers built a training pipeline they call the Kaggle-MLE Factory specifically to generate executable tasks and an interaction harness for this kind of self-directed experimentation.

What To Watch As Self-Improving Models Scale Up

The paper also reports gains on a separate test called NatureBench Lite and on agent scaffold optimization, suggesting the learned skill generalizes beyond the training benchmark.

Independent researchers will likely test whether MLExpert’s gains hold on tasks outside the Kaggle-based training data it was built on, and whether larger labs apply Windowed Interaction RL to their own frontier models in the coming months.