Research

Paper

AI LLM March 24, 2026

Off-Policy Value-Based Reinforcement Learning for Large Language Models

Authors

Peng-Yuan Wang, Ziniu Li, Tian Xu, Bohan Yang, Tian-Shuo Liu, ChenYang Wang, Xiong-Hui Chen, Yi-Chen Li, Tianyun Yang, Congliang Chen, Yang Yu

Abstract

Improving data utilization efficiency is critical for scaling reinforcement learning (RL) for long-horizon tasks where generating trajectories is expensive. However, the dominant RL methods for LLMs are largely on-policy: they update each batch of data only once, discard it, and then collect fresh samples, resulting in poor sample efficiency. In this work, we explore an alternative value-based RL framework for LLMs that naturally enables off-policy learning. We propose ReVal, a Bellman-update-based method that combines stepwise signals capturing internal consistency with trajectory-level signals derived from outcome verification. ReVal naturally supports replay-buffer-based training, allowing efficient reuse of past trajectories. Experiments on standard mathematical reasoning benchmarks show that ReVal not only converges faster but also outperforms GRPO in final performance. On DeepSeek-R1-Distill-1.5B, ReVal improves training efficiency and achieves improvement of 2.7% in AIME24 and 4.5% in out-of-domain benchmark GPQA over GRPO. These results suggest that value-based RL is a practical alternative to policy-based methods for LLM training.

Metadata

arXiv ID: 2603.23355
Provider: ARXIV
Primary Category: cs.LG
Published: 2026-03-24
Fetched: 2026-03-25 06:02

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.23355v1</id>\n    <title>Off-Policy Value-Based Reinforcement Learning for Large Language Models</title>\n    <updated>2026-03-24T15:55:02Z</updated>\n    <link href='https://arxiv.org/abs/2603.23355v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.23355v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Improving data utilization efficiency is critical for scaling reinforcement learning (RL) for long-horizon tasks where generating trajectories is expensive. However, the dominant RL methods for LLMs are largely on-policy: they update each batch of data only once, discard it, and then collect fresh samples, resulting in poor sample efficiency. In this work, we explore an alternative value-based RL framework for LLMs that naturally enables off-policy learning. We propose ReVal, a Bellman-update-based method that combines stepwise signals capturing internal consistency with trajectory-level signals derived from outcome verification. ReVal naturally supports replay-buffer-based training, allowing efficient reuse of past trajectories. Experiments on standard mathematical reasoning benchmarks show that ReVal not only converges faster but also outperforms GRPO in final performance. On DeepSeek-R1-Distill-1.5B, ReVal improves training efficiency and achieves improvement of 2.7% in AIME24 and 4.5% in out-of-domain benchmark GPQA over GRPO. These results suggest that value-based RL is a practical alternative to policy-based methods for LLM training.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.LG'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <published>2026-03-24T15:55:02Z</published>\n    <arxiv:primary_category term='cs.LG'/>\n    <author>\n      <name>Peng-Yuan Wang</name>\n    </author>\n    <author>\n      <name>Ziniu Li</name>\n    </author>\n    <author>\n      <name>Tian Xu</name>\n    </author>\n    <author>\n      <name>Bohan Yang</name>\n    </author>\n    <author>\n      <name>Tian-Shuo Liu</name>\n    </author>\n    <author>\n      <name>ChenYang Wang</name>\n    </author>\n    <author>\n      <name>Xiong-Hui Chen</name>\n    </author>\n    <author>\n      <name>Yi-Chen Li</name>\n    </author>\n    <author>\n      <name>Tianyun Yang</name>\n    </author>\n    <author>\n      <name>Congliang Chen</name>\n    </author>\n    <author>\n      <name>Yang Yu</name>\n    </author>\n  </entry>"
}