Contributor workspace
Sign in to claim tasks, verify results and earn reputation. Browsing is open to everyone.
Sign inAcross math, code and visual-reasoning benchmarks and several model families, RL with verifiable rewards improves pass@1 but the base model catches up and sometimes surpasses the RL model at large (up to 128 to 1024 samples). The authors conclude RLVR mainly sharpens the distribution within the base model's support, and that distillation from a stronger teacher, unlike RL, expands the boundary.
Preprint. Contested by later work: ProRL (preprint) reports expansion under prolonged RL with KL control and reference resets, and a question-level analysis ("Invisible Leash", preprint) reports that lost verified successes outweigh gained ones. The authors themselves note a counter-example (GRPO pass@256 on MATH500 above the base model) and that pass@k can count lucky guesses.
Verification broken into steps anyone can take. Pick an open one, do it, and submit what you found below. Matching how a check is settled earns +8; catching a failure earns +20.
No independent rerun on a different model family has been published.
ProRL reports the opposite for a 1.5B model with a different recipe; the two have not been run under one protocol.
The Reinforcement Learning committee has no members yet, so this claim can be machine-checked and verified by anyone but not committee-reviewed.
No verifications yet.
Sign in to verify.