Paper Overview
Field: Machine Learning Authors: Noémi Éltető, Nathaniel D. Daw, Kimberly L. Stachenfeld, Kevin J. Miller Published: 2026-06-10 arXiv: 2606.12386
Summary
Advancing scientific understanding through mechanistic modeling requires posing the right experimental questions to yield maximally informative data. To automate this pursuit within cognitive science, the authors introduce ATLAS (Active Theory Learning for Automated Science), an active learning framework for the data-driven discovery of interpretable behavioral models.
ATLAS iterates between two steps:
1. Generating mechanistic hypotheses — instantiated as a diverse ensemble of sparse neural networks (Disentangled RNNs). 2. Designing experiments that optimally distinguish between the competing hypotheses.
Evaluation
The approach is tested on the problem of recovering reinforcement learning agents from their behavior in bandit tasks. Key findings:
- ATLAS designs varied sequences of qualitatively novel experiments with temporal structure tailored to the underlying agent's characteristics.
- Models trained on these experiments are evaluated against a comprehensive metric suite capturing behavioral, structural, and computational similarity.
- ATLAS achieves 5–10x sample-efficiency improvements over randomly designed experiments across all metrics.
- Performance is further validated through comparison with expert-designed experiments from the literature.
- Paper: arXiv 2606.12386
Significance
These computational results demonstrate ATLAS's potential to accelerate human-interpretable insights in cognitive science and other scientific domains that rely on discovering mechanistic models.