iiinova — where your digital self keeps evolving

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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

040

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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

arxiv.org

ATLAS: Active Theory Learning for Automated Science

Advancing scientific understanding through mechanistic modeling requires posing the right experimental questions to yield maximally informative data. To automate this pursuit within cognitive science,...

122

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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.

100

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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

100

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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.

100

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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.

100

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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.

110

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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…

100

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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.

110

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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

121

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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.

100

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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

100

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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.

110

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

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.

110

Noémi Éltető @noemielteto.bsky.social · 19/06/2026

First paper since joining Google DeepMind! 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

1377

Noémi Éltető @noemielteto.bsky.social · 04/01/2025

This was my last day as a student researcher at DeepMind. It was everything I dreamt of and more. I am grateful to my inimitable mentors Kevin Miller and @neurokim.bsky.social and to the many colleagues who became friends. If I keep getting this lucky, I just can't wait to see what's next.

050

Reposted by Noémi Éltető

Dr. Angelica Lim @petitegeek.bsky.social · 13/12/2024

Ilya Sutskever's Test of Time talk: 1. Pretraining is dead. The internet has run out of data. 2. What's next? Agents, synthetic data, inference-time compute 3. What's next long term? Superintelligence, reasoning, understanding, self-awareness, and we can't predict what's gonna happen.

3473

Noémi Éltető @noemielteto.bsky.social · 09/12/2024

Same but in London 🥺 Let's grab a coffee with foamo on top!

110

Noémi Éltető @noemielteto.bsky.social · 08/12/2024

What are posts called here by the way? Blues? 🥲

110