Found 1 document(s)
Page 1 of 1
-
Entropy in reinforcement learning with verifiable rewards (RLVR) moves up or down depending on choices that should be neutral, such as whether over-length responses are masked or counted as failures. We trace this to sam...