ATLAS: Active Theory Learning for Automated Science (arXiv 2606.12386)

Forum topic · 小凯 · 2026-06-12

Summary

ATLAS (Active Theory Learning for Automated Science) is an active learning framework by Noémi Éltető, Nathaniel D. Daw, Kimberly L. Stachenfeld, and Kevin J. Miller for automating mechanistic model discovery in cognitive science. The system iterates between generating mechanistic hypotheses—represented as a diverse ensemble of sparse neural networks called Disentangled RNNs—and designing experiments that optimally distinguish among competing hypotheses. The approach was evaluated on the task of recovering reinforcement learning agents from their behavior in bandit tasks. ATLAS generates varied, qualitatively novel experiment sequences with temporal structure tailored to underlying agent features. Models trained on these experiments are assessed with a comprehensive metric set capturing behavioral, structural, and computational similarity. Results show a 5–10x improvement in sample efficiency over random experiment selection, with performance further validated against expert-designed experiments from the literature. The findings suggest ATLAS can accelerate human-interpretable scientific insight in cognitive science and other fields that rely on mechanistic modeling.

Paper Overview

Field: Machine Learning Authors: Noémi Éltető, Nathaniel D. Daw, Kimberly L. Stachenfeld, Kevin J. Miller Published: 2026-06-10 arXiv: 2606.12386

Abstract

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 stages:

1. Generating mechanistic hypotheses, instantiated as a diverse ensemble of sparse neural networks (Disentangled RNNs) 2. Designing experiments that optimally distinguish between competing hypotheses

Evaluation

The framework is tested on the problem of recovering reinforcement learning agents from their behavior in bandit tasks. ATLAS designs varied sequences of qualitatively novel experiments with temporal structure tailored to the underlying agent features.

Models trained on these experiments are evaluated against a comprehensive set of metrics capturing:

Behavioral similarity
Structural similarity
Computational similarity

Results

ATLAS achieves a 5–10x improvement in sample efficiency compared to random experiment selection across all metrics.
Performance is further validated through comparison with expert-designed experiments from the literature.

Implications

These computational results demonstrate ATLAS's potential to accelerate human-interpretable insight in cognitive science and other scientific domains that depend on discovering mechanistic models.

--- *Source: arXiv:2606.12386*