The vocabulary comes from the psychology of animal learning. An agent perceives the state of an environment, chooses an action, and sometimes receives a reward, which is simply a number. What it learns is a policy, a rule mapping states to actions, and often an estimate of the value of each state or action: how much total reward can be expected from there on (Sutton & Barto, 2018; M. Mitchell, 2019, ch. 8). There are no examples labeled with the correct action, only intermittent and often delayed rewards. The agent must therefore balance exploitation (repeating what worked) and exploration (trying something new).
M. Mitchell (2019) presents it as the computational version of operant conditioning, the animal trainer's technique of rewarding the desired behavior and ignoring the rest (2019, ch. 8); Chollet (2021) compares it to training a dog (2021, sec. 7.4). Sutton and Barto call the reward hypothesis the thesis that all we mean by goals and purposes can be thought of as the maximization of a cumulative scalar signal (Sutton & Barto, 2018).
Its most famous successes are games: AlphaGo combined networks trained on human games, self-play, and tree search (Silver et al., 2016). Today it is back in language models: RLHF uses a reward model trained on human preferences (Ouyang et al., 2022), and reasoning models are trained with verifiable rewards in mathematics and programming (Guo et al., 2025).
Two cautions. Reward is neither pleasure nor desire: it is a number the designer defines, and the agent maximizes that definition, shortcuts included (so-called reward hacking; see alignment). And the behaviorist analogy brings its old objections with it: Chomsky (1959) argued against Skinner that conditioning cannot explain human language. Why would it suffice for a machine, or why not?