One of key hallmarks of cognitive flexibility is the capability of quickly performing neural computation to accommodate dynamic, context-switching environments and deal with task uncertainty [1-3]. Accumulating evidence from human and animal experiments have suggested that the mediodorsal (MD) thalamus and prefrontal cortex (PFC) are a pair of critical partners to achieve cognitive control. e found that such multiplicative thalamocortical gating enables a local Hebbian-weight amplification for prefrontal synaptic plasticity and is the key to facilitate a low-rank structure in prefrontal computation. In the presence of outcome uncertainty (i.e., “probabilistic reward”) [5], we also found that the MD-PFC multiplicative coupling can speed up reinforcement learning (RL) with few-shot learning, improving convergence speed by 5-10 folds in context-switching tasks. During context switch, and some MD units represent post-reward prediction errors during cue-to-rule remapping or missing reward (“surprise”), supporting a Bayesian predictive coding framework in thalamocortical computation [6,7], where the MD represents contextual priors and uncertainty. The proposed multiplicative gating in a hybrid RNN-FNN interaction loop may have a broader implication in other neural circuits or behavioral tasks [8], as we have also found significantly faster (by 1-2 folds) learning speed in other RL benchmarks (such as the car-mountain and cart-pole problems in continuous control). In conclusion, our work not only suggests the role of multiplicative thalamocortical gating in enabling flexible prefrontal computation and cognitive flexibility, but also generate new insights into NeuroAI applications.