Optimizing behavioral strategy requires belief updating based on new evidence, a process that engages higher cognition. In schizophrenia, aberrant belief dynamics may lead to psychosis, but the mechanisms underlying this process are unknown, in part, due to lack of appropriate animal models and behavior readouts. Here, we address this challenge by taking two synergistic approaches. First, we generate a mouse model bearing point mutation in Grin2a (Grin2aY700X+/-), a gene that confers high-risk for schizophrenia and was recently identified by large-scale exome sequencing. Second, we develop a computationally-trackable foraging task, in which mice form and update belief-driven strategies in a dynamic environment. We found thatGrin2aY700X+/-mice perform less optimally than their wild-type (WT) littermates, showing unstable behavioral states and a slower belief update rate. Using functional ultrasound imaging, we identified the mediodorsal (MD) thalamus as hypofunctional inGrin2aY700X+/-mice, andin vivotask recordings showed that MD neurons encoded dynamic values and behavioral states in WT mice. Optogenetic inhibition of MD neurons in WT mice phenocopiedGrin2aY700X+/-mice, and enhancing MD activity rescued task deficits in Grin2aY700X+/-mice. Together, our study identifies the MD thalamus as a key node for schizophrenia-relevant cognitive dysfunction, and a potential target for future therapeutics.