Thread by @EltetoNoemi

First paper since joining @GoogleDeepmind! We present 🌍ATLAS (Active Theory Learning for Automated Science), a pipeline that generates interpretable mechanistic models from data and optimizes experiments to test them.

Thread below

2/14 Uncovering the algorithms behind behaviour is a central goal of cognitive science. This usually requires hard-won datasets and careful experiment design. Active learning provides a framework for how to collect maximally informative data efficiently.

3/14 ATLAS automates the scientific method in a closed loop.

1. It optimizes experiments by maximizing disagreement between competing hypotheses.

2. It generates novel mechanistic hypotheses, consistent with the dataset so far, to drive further experimentation.

4/14 The ATLAS loop is started either with a small initial dataset or user-provided hypotheses.

5/14 First, let’s see some ATLAS-designed experiments for user-provided hypotheses!

Below, the hypotheses are manually implemented Q-learning agents that differ in their learning rates. Experiments are binary matrices determining which actions will be rewarded at which timestep.

6/14 ATLAS designs experiments with intricate temporal structures—like overlapping reward blocks—which force competing models to maximally reveal their differences (in terms of p(action=action 1). Notice how the block durations are precisely tailored to the hypotheses at hand!

7/14 But active learning has a caveat: the hypothesised model must be specified correctly. If the question is wrong, the answer may lead us further astray.

8/14 For example, if a model lacks the parameters to capture a Q-learning agent's true action values, it doesn't just fail to explain existing data but it also misinforms how we collect future data…

9/14 The ATLAS loop—alternating between data-driven hypothesis generation and experiment design—is designed to take us from weak models to strong ones using as few experiments as possible.

10/14 We tasked ATLAS with recovering RL agents (like Q-learning) from their behaviour in reward learning tasks. By building a diverse curriculum of temporally structured experiments, ATLAS uncovers a neural network whose computational graph is isomorphic to the ground truth.

11/14 Compared with random experimentation, ATLAS achieved a 5–10x improvement in sample efficiency, on three criteria of mechanistic modeling, and for two example agents.

12/14 More impressively, ATLAS matches or even beats expert-designed experiments from the literature: parametric bandits with slowly drifting reward probabilities.

13/14 By tailoring its experiments dynamically to the specific agent being studied, ATLAS consistently recovers the true computational structure and dynamics in a fraction of the time.

14/14 By linking data-driven hypothesis generation with hypothesis-driven experiment selection, ATLAS shows real potential to accelerate the discovery of interpretable insights in cognitive science and beyond.

Read the full paper here: arxiv.org/abs/2606.12386

14+1/14 It’s been an honour and joy to work on this project with @nathanieldaw, @neuro_kim, and @kevinjmiller10 at the Neuroscience Lab @GoogleDeepMind.