Trajectory Bellman Residual Minimization: A Simple Value-Based Method for LLM Reasoning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Yuan, Yurun, Chen, Fan, Jia, Zeyu, Rakhlin, Alexander, Xie, Tengyang
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912702857740288
author Yuan, Yurun
Chen, Fan
Jia, Zeyu
Rakhlin, Alexander
Xie, Tengyang
author_facet Yuan, Yurun
Chen, Fan
Jia, Zeyu
Rakhlin, Alexander
Xie, Tengyang
contents Policy-based methods currently dominate reinforcement learning (RL) pipelines for large language model (LLM) reasoning, leaving value-based approaches largely unexplored. We revisit the classical paradigm of Bellman Residual Minimization and introduce Trajectory Bellman Residual Minimization (TBRM), an algorithm that naturally adapts this idea to LLMs, yielding a simple yet effective off-policy algorithm that optimizes a single trajectory-level Bellman objective using the model's own logits as $Q$-values. TBRM removes the need for critics, importance-sampling ratios, or clipping, and operates with only one rollout per prompt. We prove convergence to the near-optimal KL-regularized policy from arbitrary off-policy data via an improved change-of-trajectory-measure analysis. Experiments on standard mathematical-reasoning benchmarks show that TBRM consistently outperforms policy-based baselines, like PPO and GRPO, with comparable or lower computational and memory overhead. Our results indicate that value-based RL might be a principled and efficient alternative for enhancing reasoning capabilities in LLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2505_15311
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Trajectory Bellman Residual Minimization: A Simple Value-Based Method for LLM Reasoning
Yuan, Yurun
Chen, Fan
Jia, Zeyu
Rakhlin, Alexander
Xie, Tengyang
Machine Learning
Artificial Intelligence
Computation and Language
Policy-based methods currently dominate reinforcement learning (RL) pipelines for large language model (LLM) reasoning, leaving value-based approaches largely unexplored. We revisit the classical paradigm of Bellman Residual Minimization and introduce Trajectory Bellman Residual Minimization (TBRM), an algorithm that naturally adapts this idea to LLMs, yielding a simple yet effective off-policy algorithm that optimizes a single trajectory-level Bellman objective using the model's own logits as $Q$-values. TBRM removes the need for critics, importance-sampling ratios, or clipping, and operates with only one rollout per prompt. We prove convergence to the near-optimal KL-regularized policy from arbitrary off-policy data via an improved change-of-trajectory-measure analysis. Experiments on standard mathematical-reasoning benchmarks show that TBRM consistently outperforms policy-based baselines, like PPO and GRPO, with comparable or lower computational and memory overhead. Our results indicate that value-based RL might be a principled and efficient alternative for enhancing reasoning capabilities in LLMs.
title Trajectory Bellman Residual Minimization: A Simple Value-Based Method for LLM Reasoning
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2505.15311