Found 1 document(s)
Page 1 of 1
-
Group relative policy optimization learns from a prompt only when the rewards in the sampled group disagree. With a binary verifier, a group that is entirely correct or entirely incorrect has within-group variance zero a...