Poster at CCN 2026

Abstract

People often develop conventions that allow them to divide cognitive labor efficiently, even without real-time reward signals to guide coordination. The computational dynamics underlying this process remain poorly understood. We apply SPICE (Sparse and Interpretable Cognitive Equations), a novel framework that distills recurrent neural networks into symbolic equations, to model convention formation in a collaborative visual working memory task. Participants (N = 128 dyads) worked in pairs to memorize a 4x4 grid. Each participant could see the square their partner was studying in real time, allowing them to form tacit conventions to split up the grid. We applied SPICE to model participants’ memory encoding strategies as a latent value function that evolves over time. The dynamic equations discovered by SPICE did not include terms for how long it had been since the participant or their partner had studied a stimulus, suggesting that participants’ encoding strategies were not sensitive to memory decay. Instead, we found evidence for a simple repulsion heuristic – Participants avoided encoding stimuli that their partner had already encoded.