142

Putting the Value Back in RL: Better Test-Time Scaling by Unifying LLM Reasoners With Verifiers

Main:9 Pages
9 Figures
Bibliography:4 Pages
Appendix:3 Pages
Abstract

Prevalent reinforcement learning~(RL) methods for fine-tuning LLM reasoners, such as GRPO or Leave-one-out PPO, abandon the learned value function in favor of empirically estimated returns. This hinders test-time compute scaling that relies on using the value-function for verification. In this work, we propose RLV^V that augments any ``value-free'' RL method by jointly training the LLM as both a reasoner and a generative verifier using RL-generated data, adding verification capabilities without significant overhead. Empirically, RLV^V boosts MATH accuracy by over 20\% with parallel sampling and enables 832×8-32\times efficient test-time compute scaling compared to the base RL method. RLV^V also exhibits strong generalization capabilities for both easy-to-hard and out-of-domain tasks. Furthermore, RLV^V achieves 1.21.6×1.2-1.6\times higher performance when jointly scaling parallel and sequential test-time compute with a long reasoning R1 model.

View on arXiv
Comments on this paper