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

TL;DR

This paper introduces RL$^V$, a method that enhances RL fine-tuning of LLM reasoners by enabling joint training for reasoning and verification, significantly improving test-time scaling and accuracy.

Contribution

RL$^V$ is a novel approach that unifies reasoning and verification training, boosting performance and enabling efficient test-time compute scaling without major overhead.

Findings

RL$^V$ increases MATH accuracy by over 20% with parallel sampling.

Enables 8-32× more efficient test-time compute scaling.

Achieves 1.2-1.6× higher performance with combined parallel and sequential scaling.

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. Yet if parallel test-time compute is already part of the deployment plan, training should be designed to support it. In this work, we propose RL$^V$ 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, RL$^V$ boosts MATH accuracy by over 20\% with parallel sampling and enables $8-32\times$ efficient test-time compute scaling compared to the base RL method. RL$^V$ also exhibits strong generalization capabilities for both easy-to-hard and out-of-domain tasks. Furthermore, RL$^V$ achieves $1.2-1.6\times$ higher performance when jointly scaling parallel and sequential test-time compute with a long reasoning R1 model. More broadly, RL$^V$ instantiates the principle of co-training for test-time scaling: jointly optimizing for task performance and a capability useful at inference, using data that RL training already produces.