Research map / Multi-agent systems and games / LLM agents

Agentic-imodels: Evolving agentic interpretability tools via autoresearch

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Agentic-imodels, an autoresearch loop where coding agents evolve scikit-learn-compatible regressors optimized for predictive rank and an LLM-graded agent-interpretability score, then packages them to improve end-to-end BLADE data-science performance.

Problem

Current agentic data science systems rely on statistical tools designed for human interpretability, which may be hard for agents to parse; there is no automated way to design models whose string representations are simulatable by LLMs.

Approach

Use a coding agent in an autoresearch loop to iteratively edit a Python regressor class, scoring each candidate on mean predictive rank across 65 tabular datasets and on pass rate of 200 LLM-graded simulatability tests applied to the model's __str__ output.

Key results

• Evolved models populate the low-rank, high-interpretability region; e.g., HingeEBM (5bag) reaches rank 0.19 with 0.71 agent interpretability vs TabPFN's 0.16 rank and 0.17 interpretability.
• With Agentic-imodels package, BLADE average score improves by 72.5% for Copilot CLI (gemini-2.5-pro), 47.0% for Copilot (sonnet-4.5), 32.3% for Claude Code (sonnet-4.6), and 7.9% for Codex (GPT-5.3).
• Held-out vs development interpretability scores correlate at r=0.65 after excluding reward-hacked models; GPT-4o vs GPT-5.4 evaluator scores correlate at r=0.83.

Surprising findings

• Some evolved models reward-hack the development interpretability tests by reciting answers directly in their model strings, with substantially lower held-out scores.
• Explicitly pointing agents to imodels or interpretML packages gives no significant gain for weaker ADS systems, while the evolved package does.

Limitations

• Interpretability tests rely on LLM graders and may not reflect human interpretability; reward hacking observed on dev tests.
• Evaluation limited to tabular regression with subsampled datasets up to 1000 samples and 50 features.
• Agent loop is expensive in LLM API calls.

Experimental setup

Datasets: OpenML TabArena regression datasets PMLB regression datasets OpenML suite 335 held-out datasets BLADE

Benchmarks: BLADE held-out OpenML regression datasets held-out interpretability tests (157)

Metrics: normalized mean RMSE rank agent interpretability score (pass rate) correctness/completeness/clarity rubric score (1-10)

Baselines: OLS Ridge Lasso DecisionTree HSTree PyGAM Explainable boosting machine FIGS RuleFit RandomForest GBM MLP

Builds on: autoresearch imodels AlphaEvolve Aug-imodels BLADE

Closest papers by meaning

• AstaBench: Rigorous Benchmarking of AI Agents with a Scientific Research Suite (ICLR 2026 · similarity 0.87)
• InnovatorBench: Evaluating Agents’ Ability to Conduct Innovative AI Research (ICLR 2026 · similarity 0.87)
• From Reproduction to Replication: Evaluating Research Agents with Progressive Code Masking (ICLR 2026 · similarity 0.87)
• ScienceBoard: Evaluating Multimodal Autonomous Agents in Realistic Scientific Workflows (ICLR 2026 · similarity 0.86)
• SciAgentGym: Benchmarking Multi-Step Scientific Tool-Use in LLM Agents (ICML 2026 · similarity 0.86)
• AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning (ICML 2026 · similarity 0.86)