Definition
Plain language
A reward signal that's just yes or no — did it work or didn't it.
As stated in the literature
A scalar training signal taking values in {0,1}, typically used in RLVR pipelines where verifiable answer correctness is the only learning signal.
Also called: binary rewards
Why it matters: Binary rewards are simple and unhackable when the verifier is solid, but they make learning hard when most attempts either all succeed or all fail.
For example, in RL training on math problems, the model gets 1 if the final answer matches the ground truth and 0 otherwise, with no partial credit.
Heard on the show
“The naive way: give it a binary reward at the end of each task, train.”Episode 008 — Why Long-Horizon AI Agents Get Stuck, and a Milestone-Based Fix That Helps