What Outcome RL Buys: Finishing, Reliability, Rarely Reach

Shan Yang

arXiv:2610.34894 · 2026-10-28 공개 · arXiv · PDF

token-budget reliability pass1 physics-problems math-problems thinking-models outcome-rl finishing

Abstract

Outcome RL raises pass@1 but not pass@k at large k (Yue et al., 2025). But every pass@1 is scored under a token budget. Every recipe also trains under one. On physics and math problems at fixed token budgets, we compare the pass@1 of 20 released RL models (1.5B to 32B) with that of the thinking models they started from. We ask whether each pass@1 gain is changed behaviour or new knowledge. We divide each change in pass@1 into three parts. Finishing measures answering before the budget runs out. Reliability measures being right once finished. Reach measures problems RL solves that its base answered wrongly in all eight tries, minus problems the base solves that RL answered wrongly in all eight. We find four things. (1) Where the base often hits the token budget before answering, most of the change, gain or loss, is finishing (82% at 4k and 8k). Given more tokens, the base closes most of the finishing gap itself. (2) Where the base has enough tokens, the gain is reliability: the RL model is graded right more often, even on problems both finish. (3) By our definition, RL rarely solves what its base cannot. Reach adds at most 2.1 percentage points on any of our four physics and math sets. We checked 144 physics problems that only the RL model solved. Of these, 3.5% hold up when resampled or read. In a random sample, nine of ten new RL steps a human confirmed also appear in the base's samples. (4) In a small-scale RL run on an 8B model, the gain on new problems is finishing, with reliability flat. Supplying each problem's physical setup was not shown to improve on plain RL.

한국어 요약

한 줄 요약

Outcome RL은 pass@1 성능을 높이지만, 대부분은 토큰 예산 내 완료율 개선이며, 새로운 지식은 거의 생성되지 않는다.

핵심 기여도

핵심 아이디어

Outcome RL은 최종 정답만 보상하는 방식으로, pass@1 성능을 높인다. 그러나 이 성능 개선이 새로운 지식의 학습인지, 행동 변화인지 구분이 어려웠다. 본 연구는 pass@1의 변화를 세 가지 요인으로 분리했다: **Finishing**(예산 내 완료), **Reliability**(정확도), **Reach**(기반 모델이 틀린 문제를 RL이 맞춘 비율). 이 중, **Finishing**이 토큰 예산이 제한적일 때 대부분의 성능 차이를 차지하며, 예산이 충분하면 **Reliability**가 주요 요인이 된다. **Reach**는 매우 제한적이며, RL이 기반 모델이 풀지 못한 문제를 해결하는 경우는 드물다. 이는 Outcome RL이 새로운 지식을 학습하는 것이 아니라, 기존 지식을 더 효율적으로 활용하는 방식으로 성능을 개선한다는 통찰을 제공한다.

기술적 접근법

주요 결과

의의 및 한계

실용적 활용