Results — MLE-Bench Low

We evaluated Qualia, our autonomous research agent, on MLE-Bench Low, a set of 22 Kaggle machine learning competitions that test an agent's ability to train models, engineer features, and produce competition-grade submissions. Our agent performed these with no human intervention other than the original prompt, spawning in many cases additional agents to help with testing out multiple modeling and feature engineering choices in parallel (see the Agents column), or to build on its prior attempts. Click on any row to see the chat history and all agents spawned, as well as the tasks into which the agent split the work. Qualia medalled on 77.2% of the tasks, tying for 4th place on the dataset while using far less time than the allotted maximum (24h).