Agentic-imodels

Evolving agentic interpretability tools via autoresearch

Quick start • Repo layout • Discovered models • How the loop works • Paper

Plugin skill: To use the developed models in your own data-science projects, just add a pointer to the skill file at in your CLAUDE.md / AGENTS.md.

We built a library of 10 `scikit-learn`-compatible regressors whose string representations are explicitly optimized to be read by another LLM — interpretable *by agents*, not just by humans. We did this by using coding agents with a fixed evaluation harness and a single Python file. that is optimized for:

• Predictive performance - root-mean squared error (RMSE) across many datasets
• Agent Interpretability — fraction of LLM-graded tests passed

Quickstart

**Requirements:** Python 3.10+ and [uv](https://docs.astral.sh/uv/).

git clone https://github.com/csinva/imodels-evolve
cd imodels-evolve
uv sync

Use the curated discovered models

from sklearn.datasets import fetch_california_housing
from sklearn.model_selection import train_test_split
from agentic_imodels import HingeEBMRegressor

X, y = fetch_california_housing(return_X_y=True)
X_tr, X_te, y_tr, y_te = train_test_split(X, y, random_state=0)

model = HingeEBMRegressor()
model.fit(X_tr, y_tr)

print(model)               # human/LLM-readable equation
preds = model.predict(X_te)

Ever