Social Values and Reinforcement Learning

    Professor Norberto M. Grzywacz presented intriguing research on social influence models at Loyola University Chicago this semester (Fall 2025). His study focused on two types of computational methods used to model personal values and how these values can influence decision-making. The first is the standard model in the field, the Agent-Based Social-Influence Model. In this model, values are defined by the distance between the values that two individuals hold. For instance, if individuals hold similar values, they will attract one another. Conversely, if they hold opposite values, they will repel one another. If the distance between values is moderate, they will attract one another, but to a lesser extent. In Dr. Grzywacz’s research, he proposes a second computational model—the Environmental-Social model—which considers how values may be influenced by reinforcement learning and other social rules. The main findings were that the Environmental-Social model was more accurate in predicting group formation based on values, the extent of polarization, and the ways in which groups are complex entities that can become unstable and vary in their values.     

    A study that relates to Dr. Grzywacz’s research is “Neural responses to social rejection reflect dissociable learning about relational value and reward” by Dr. Begüm G. Babür and colleagues (2024). The researchers investigated social learning, specifically whether an outcome was rewarding and how these social cues can update beliefs and potentially alter relationships. Participants first completed a questionnaire on their trustworthiness, and a week later played a computer-based Trust Game in which they could send points worth money to different players (these players were pre-programmed by the computer). Participants also had to choose their partner in this game, and the feedback they received included whether they were matched with the other player and how highly the other player ranked them. Functional MRI (fMRI) was used to record brain responses during the Trust Game. The researchers used Kullback–Leibler (KL) divergence, a type of computational model, to quantify fMRI signal changes based on the reward feedback the participants received. They found that the orbitofrontal cortex was a crucial region of activation during social learning and that the highest reward occurred when participants were matched with another player and were ranked highly by that player. 

    Connecting this study to Dr, Grzywacz’s research, it seems that social values (in this case, honesty) play a significant role in participants’ ratings of other players and in their reward-based decisions. This is supported by Dr. Babür and colleagues (2024), who observed that players with more similar voxel patterns in fMRI results were rated similarly during the Trust Game. These results align with the standard Social-Influence Model, which states that individuals with similar values tend to attract one another. This study also strongly supports Dr. Grzywacz’s Environmental-Social model, as participants updated their decisions based on the outcomes that produced the highest reward. In this experiment, this meant choosing partners who they ended up matching with and who they received high ratings from. In other words, participants felt most rewarded when the other person reciprocated social value. Overall, it seems that decision-making can be influenced by a variety of factors, especially those involving social values, reward, and rejection. By understanding the neurological components and computational methods used to model them, social cognition and relationships may be better understood. 

B.G. Babür, Y.C. Leong, C.X. Pan, & L.M. Hackel, Neural responses to social rejection reflect dissociable learning about relational value and reward, Proc. Natl. Acad. Sci. U.S.A. 121 (49) e2400022121, https://doi.org/10.1073/pnas.2400022121 (2024).