Most theories of social exchange distinguish between two different types of cooperation, depending on whether or not cooperation occurs conditional upon the partner's previous behaviors. Here, we used a multinomial processing tree model to distinguish between positive and negative reciprocity and cooperation bias in a sequential Prisoner's Dilemma game. In Experiments 1 and 2, the facial expressions of the partners were varied to manipulate cooperation bias. In Experiment 3, an extinction instruction was used to manipulate reciprocity. The results confirm that people show a stronger cooperation bias when interacting with smiling compared to angry-looking partners, supporting the notion that a smiling facial expression in comparison to an angry facial expression helps to construe a situation as cooperative rather than competitive. Reciprocity was enhanced for appearance-incongruent behaviors, but only when participants were encouraged to form expectations about the partners' future behaviors. Negative reciprocity was not stronger than positive reciprocity, regardless of whether expectations were manipulated or not. Experiment 3 suggests that people are able to ignore previous episodes of cheating as well as previous episodes of cooperation if these turn out to be irrelevant for predicting a partner's future behavior. The results provide important insights into the mechanisms of social cooperation.