#AGENTIC-IMODELS: Evolving #datascience Tools for #agenticai Interpretability

Standard data science tools were designed for human eyes—using complex visualizations and diagrams. But for AI agents, these tools are a "mismatch" that leads to unreliable analysis. In this video, we dive into AGENTIC-IMODELS, a groundbreaking research project from Microsoft Research and NUS that introduces an "autoresearch loop" where AI agents evolve their own machine learning tools. By optimizing for a novel Agent Interpretability Score, these agents have developed a library of models—like HingeEBM and SmartAdditive—that are specifically designed to be "read" and simulated by LLMs. What we cover: The Problem: Why human-centric interpretability tools derail AI Data Science (ADS) systems. The Solution: The autoresearch loop that prompts coding agents to iteratively build scikit-learn-compatible models. Simulatability: How 200 LLM-graded tests measure if an agent can truly understand a model's logic through text alone. The Results: How these evolved models improved downstream performance for GitHub Copilot, Claude Code, and Codex by up to 73% on the BLADE benchmark. Model Deep Dive: How HingeEBM and SmartAdditive use "display-predict decoupling" and Laplacian smoothing to balance accuracy with clarity. Resources: GitHub: github.com/csinva/agentic-imodels Paper: "AGENTIC-IMODELS: Evolving agentic interpretability tools via autoresearch"