· 21 min read
RLVR is blind to redundancy
Many tasks have more than one minimal answer, none containing another. A reward that scores each rollout on its own cannot tell a new answer from a redundant copy of one it already has.
· 21 min read
Many tasks have more than one minimal answer, none containing another. A reward that scores each rollout on its own cannot tell a new answer from a redundant copy of one it already has.