Definition
Plain language
A mathematical equivalence showing that training an AI with rewards is, underneath, the same operation as updating a belief with new evidence.
As stated in the literature
The known correspondence between KL-regularized reinforcement learning and Bayesian inference, in which the optimal leashed policy is exactly a reward-tilted posterior over the base model's outputs; exploited to replace gradient-based training with sampling from that posterior.
Also called: RL as inference
Why it matters: This equivalence lets researchers swap costly gradient training for sampling, opening alternative and sometimes cheaper ways to steer a model.
For example, it shows that nudging a model toward higher-reward outputs is mathematically the same as updating your beliefs about which of its outputs are good.