Thalamic regulation of reinforcement learning strategies across prefrontal-striatal networks.

Human decision-making involves model-free and model-based reinforcement learning (RL) strategies, largely implemented by prefrontal-striatal circuits. Combining human brain imaging with neural network modelling in a probabilistic reversal learning task, we identify a unique role for the mediodorsal thalamus (MD) in arbitrating between RL strategies. While both dorsal PFC and the striatum support rule switching, the former does so when subjects predominantly adopt model-based strategy, and the latter model-free. The lateral and medial subdivisions of MD likewise engage these modes, each with distinct PFC connectivity. Notably, prefrontal transthalamic processing increases during the shift from stable rule use to model-based updating, with model-free updates at intermediate levels. Our CogLinks model shows that model-free strategies emerge when prefrontal-thalamic mechanisms for context inference fail, resulting in a slower overwriting of prefrontal strategy representations - a phenomenon we empirically validate with fMRI decoding analysis. These findings reveal how prefrontal transthalamic pathways implement flexible RL-based cognition.