Introduction Volatile interpersonal relationships are a core feature of Borderline Personality Disorder (BPD), and lead to devastating disruption of patients’ personal and professional lives. Quantitative models of social decision making and learning hold promise for defining the underlying mechanisms of this problem.

Methods In this study, women with BPD and healthy controls interacted with a confederate, then played an extended reward learning task that requires learning about non-social and social cue reward probability (The Social Valuation Task). Subject distress after debriefing was measured using a language marker (self-referentiality). Weighting of social and non-social cues in subject decisions was tested in mixed-effects regression models. Changes in learning rates in response to volatility for the non-social and social rewards was tested in a Rescorla-Wagner learning model.

Results BPD but not control subject language had fewer distress markers after debriefing (including disclosure of deception). Compared to controls, BPD subjects weighted social cues more heavily, but they showed significantly blunted learning responses to reward volatility for both the non-social and social cues.

Conclusions This is the first report of patient behavior in the Social Valuation Task. The results suggest that BPD subjects expect higher social and non-social volatility than do controls. These findings lay the groundwork for a neuro-computational dissection of social and non-social belief updating in BPD. Such an understanding, bridging circuity with clinical phenomenology through computation, portends clinical interventions that more directly target pathophysiology.