reward hypothesis

The reward hypothesis is RL's founding bet: that every goal we care about can be captured as maximizing the expected total of a single scalar reward. Win the game, fold the protein, keep the room at twenty-one degrees, hold a helpful and honest conversation—each, the hypothesis says, can be phrased as "earn as much cumulative reward as possible." It's a bold claim that one number can encode purpose.

This idea is what lets RL be a single, general framework rather than a bag of tricks. But it's a hypothesis, not a theorem, and it bites in practice: squeezing a rich, multi-faceted goal into one number invites reward hacking, where the agent maximizes the letter of the reward while violating its spirit. Whether everything truly reduces to scalar reward remains genuinely debated.