Reinforcement Learning Constructs Shape Subjective Connection and Social Memory.

Courtney A. Jimenez & Meghan L. Meyer · Department of Psychology, Columbia University · PsyArXiv, 2026 (Under Review)

What it’s about: Modeling how reward, expected value, and prediction error — core reinforcement learning constructs — shape people’s moment-to-moment sense of social connection and what they remember from social interactions.

Approach: Participants viewed a target character’s behavior for social reward across a continuous timecourse (using reality-TV video clips), then completed a social decision-making task and a surprise memory test. A reinforcement learning model (reward prediction error updating expected value, softmax choice rule) was fit to participants’ friendship decisions in R.

Results:

  • Friendship reward value, expected value, and prediction error all significantly predicted participants’ subjective connection ratings
  • Individual differences in fitted RL parameters (learning rate α, inverse temperature β) related to participants’ social experience
  • Memory was better for moments of lower subjective connection and higher social reward value — suggesting affectively surprising or motivationally salient moments are prioritized in memory

The Dorsomedial Prefrontal Cortex Prioritizes Social Learning During Rest.

Courtney Jimenez & Meghan Meyer · Department of Psychology, Columbia University · PNAS, 2024

What it’s about: Testing whether the brain preferentially consolidates social information during rest, using neural pattern reinstatement — directly relevant to how memory systems prioritize different types of information.

Approach: Participants encoded social and non-social stimuli, rested, and completed a memory test. High-dimensional fMRI pattern reinstatement during rest was measured in the dorsomedial prefrontal cortex (DMPFC, a social-cognition hub) and the ventrolateral prefrontal cortex (VLPFC, implicated in non-social memory), and related to later memory performance across early, middle, and late rest periods.

Results:

  • Social memory performance was positively correlated with DMPFC reinstatement during rest; non-social memory performance was positively correlated with VLPFC reinstatement
  • Social consolidation in the DMPFC was specific to early rest, while non-social consolidation in the VLPFC showed no such early-rest specificity
  • Findings support the DMPFC — a key default-network node — as prioritizing social information consolidation, consistent with the brain being “social by default”