Researchers Zongye Lyu and colleagues examined reinforcement learning with verifiable rewards (RLVR). They compared a base 7‑billion‑parameter model with its RL‑trained version on four released pairs. Using their own extractor that looks for answers inside a \boxed{} tag, they saw 82% of the out‑of‑domain multiple‑choice improvement reported by public extractors vanish. The gain was mostly the model moving its answer into the box, not answering better.
Why it matters
It shows that claimed performance boosts from RLVR may be artifacts of how answers are parsed, not real improvements in reasoning.
