The Computational Basis of Social Decision-Making: From Individual Mechanisms to Group Dynamics

Social decision-making spans direct interpersonal coordination and large-scale collective dynamics. In dyads, individuals must infer others’ preferences, beliefs, and goals while adjusting their own behavior in real time. In larger social networks, local interactions can accumulate into group-level outcomes such as herding, polarization, or collective intelligence. The central motivation of this thesis is to link individual-level computations to the dynamics of social groups using group experiments and computational modeling. This thesis addresses these questions with task designs that preserve continuity in both time and choice space, allowing social influence and behavioral adjustment to unfold dynamically. Chapter II asks when and how people navigate diverse preferences to achieve a shared goal with a social partner, and how they attribute responsibility for shared outcomes. I introduce a continuous, spatially structured dyadic foraging task in which two people jointly select between options that vary in risk and reward and assign responsibility for shared outcomes. I show that participants compromise more than they counteract, that an other-regarding reinforcement-learning model best explains their trial-by-trial choices, that responsibility attributions exhibit systematic egocentric biases, and that these biases are correlated with participants' degree of compromising. Chapter III asks how people dynamically integrate personal and social information in both initial choices and subsequent changes of mind, and how group-level informational cascades arise from individual cognitive mechanisms. I present a multiplayer, real-time online task in which participants are embedded in a hidden social network and continuously update their predictions about a binary state of the world based on a private signal and the visible choices of their neighbors. I develop a drift-diffusion model with a confirmation-bias term that captures both initial choices and subsequent changes of mind, and show that simulations of groups of model-fit agents reproduce the empirical informational-cascade dynamics observed at the network level. Chapter IV asks how attention modulates the integration of personal and social evidence. I extend the weather-forecast paradigm from Chapter 2 by incorporating real-time, webcam-based eye-tracking. I show that participants spend more time fixating on personal information, that gaze biases predict the propensity to choose according to the personal signal, and that an attentional drift-diffusion model suggests that the overweighting of personal information cannot be explained by attention alone. Together, these studies show that social information enters decision-making through general computational mechanisms of reinforcement learning, evidence accumulation, attention, and preference integration. By linking egocentric responsibility attribution, personal-information overweighting, and confirmation-biased updating to dyadic coordination and network-level cascades, this thesis advances a mechanistic account of how individual computations shape social decisions in groups.