We investigate theoretically and experimentally the crisis bargaining model, a dynamic game of two-sided incomplete information with player types drawn from a commonly known distribution. Little work has been done to analyze whether and how players update their beliefs in such games.Within the experiment we elicited beliefs from players about their opponent's type using a proper scoring rule. We implement two treatments that vary the cost of backing down to the first mover after initial entry, generating sharp comparative static predictions in both beliefs and strategies. We find that players do update their beliefs in the predicted directions after observing some of the action choices. However, we highlight evidence of conservative updating relative to rational expectations.