June 2025

Discovering cognitive strategies with tiny recurrent neural networks

Introduces a modeling approach using very small (1-4 units) recurrent neural networks (RNNs) to discover interpretable cognitive strategies in animal and human reward-learning tasks.
Tiny RNNs outperform classical cognitive models in predicting choices, revealing novel behavioral patterns (e.g., variable learning rates, state-dependent perseveration) that are missed by standard models.
The trained RNNs are interpretable through dynamical systems analysis, enabling visualization and comparison with classic cognitive theories.
The framework allows dimensionality estimation of behavior and provides insights into both biological cognition and AI meta-reinforcement learning.

Examines how neural dynamics during social interaction can be partitioned into “shared” and “unique” subspaces in both mice and artificial agents.
Finds GABAergic neurons in the dorsomedial prefrontal cortex (dmPFC) display significantly greater shared neural dynamics across brains than glutamatergic neurons, especially during social behavior.
Disruption of shared neural subspace components impairs social interactions in both biological and artificial systems.
Reveals that shared neural dynamics are a generalizable feature driving social behavior in interacting agents (biological or AI).

Introduces “Centaur,” a large foundation model (fine-tuned Llama 3) trained on Psych-101, a massive dataset of trial-level human cognitive/behavioral data from 160 experiments.
Centaur predicts human choices better than both base LLMs and domain-specific cognitive models, even in new tasks and domains, and its representations align more closely with human brain activity after tuning.
Demonstrates strong generalization: can simulate human-like behavior with only natural language descriptions of tasks.
Provides a platform for testing, falsifying, and refining cognitive theories at scale.