Agentic-imodels: Evolving agentic interpretability tools via autoresearch

2026-05-05

AI Summary

This work addresses the mismatch between existing human-oriented interpretability tools and the cognitive capabilities of autonomous agents in data science systems. We propose a novel agent-centered interpretability paradigm that, for the first time, employs a large language model (LLM) as an interpretability evaluator, assessing the understandability of model string representations through “simulatability.” By integrating an evolutionary algorithm with this LLM-driven interpretability metric, we establish an autonomous research loop that jointly optimizes predictive performance and agent interpretability, yielding scikit-learn-compatible regressors. The evolved models demonstrate strong generalization on unseen datasets and benchmarks, achieving up to a 73% improvement in end-to-end performance for AI coding assistants—including Copilot CLI, Claude Code, and Codex—on the BLADE benchmark.

Abstract

Agentic data science (ADS) systems are rapidly improving their capability to autonomously analyze, fit, and interpret data, potentially moving towards a future where agents conduct the vast majority of data-science work. However, current ADS systems use statistical tools designed to be interpretable by humans, rather than interpretable by agents. To address this, we introduce Agentic-imodels, an agentic autoresearch loop that evolves data-science tools designed to be interpretable by agents. Specifically, it develops a library of scikit-learn-compatible regressors for tabular data that are optimized for both predictive performance and a novel LLM-based interpretability metric. The metric measures a suite of LLM-graded tests that probe whether a fitted model's string representation is "simulatable" by an LLM, i.e. whether the LLM can answer questions about the model's behavior by reading its string output alone. We find that the evolved models jointly improve predictive performance and agent-facing interpretability, generalizing to new datasets and new interpretability tests. Furthermore, these evolved models improve downstream end-to-end ADS, increasing performance for Copilot CLI, Claude Code, and Codex on the BLADE benchmark by up to 73%

Problem

• agentic interpretability
• autonomous data science
• LLM simulatability
• agent-facing interpretability
• interpretable machine learning

Innovation

• agentic interpretability
• autoresearch
• LLM-simulatable models
• agent-facing interpretability
• evolved regressors