A key behavioral characteristic of autistic people are difficulties with mentalization-understanding and reacting to the mental states of others. Despite their social and clinical implications, the specific neuro-cognitive mechanisms that may give rise to these difficulties have remained elusive. In my presentation, I will argue that one reason for this impasse may be limitations of the experimental approaches used to assess metalization. I will present a novel behavioral paradigm and computational model that can specifically assess adaptive mentalization. Using this paradigm and model in fMRI experiments, we have identified a multivariate neural fingerprint that can predict out of sample to what degree people can dynamically adapt their mental models of others to their changing strategies. We have also used this model and neural fingerprint in a preregistered study of autistic individuals, revealing that stronger autistic traits are associated with both a decreased ability for adaptive mentalization and weaker expression of the very same multivariate neural fingerprint. We hope that these results illustrate a specific deficit in adaptive mentalization that is linked to autistic traits.