Agentic-imodels: Evolving interpretability tools for AI agents via autoresearch

A new arXiv preprint introduces Agentic-imodels, an agentic autoresearch loop that evolves data-science tools optimized for interpretability by AI agents rather than humans. The system develops a library of scikit-learn-compatible regressors for tabular data, balancing predictive performance with a novel LLM-based interpretability metric. This metric uses LLM-graded tests to assess whether a model's string representation is 'simulatable' by an LLM, meaning the LLM can answer questions about the model's behavior solely from its string output. The research aims to address the gap where current agentic data science systems rely on human-interpretable tools, which may not be optimal for agent-based analysis. The paper is available on arXiv under identifier 2605.03808.