Value-Guided Search for Efficient Chain-of-Thought Reasoning

Kaiwen Wang, Jin Peng Zhou, Jonathan Daniel Chang, Zhaolin Gao, Nathan Kallus, Kianté Brantley, Wen Sun

Keywords: Value-Guided Search, Chain-of-Thought Reasoning, Competition Math

Abstract:

In this paper, we propose a simple and efficient method for value model training on long-context reasoning traces. Compared to existing process reward models (PRMs), our method does not require a fine-grained notion of ``step,'' which is difficult to define for long-context reasoning models.
By collecting a dataset of 2.5 million reasoning traces, we train a 1.5B token-level value model and apply it to DeepSeek models for improved performance with test-time compute scaling. We find that block-wise value-guided search (\texttt{VGS}) with a final weighted majority vote achieves better test-time scaling than standard methods such as majority voting or best-of-$n$.
Moreover, \texttt{VGS} significantly reduces the inference FLOPs required to achieve the same performance of majority voting. Our dataset, model and codebase are open-sourced at \codeurl.

Primary Area: Reinforcement learning (e.g., decision and control, planning, hierarchical RL, robotics)

Submission Number: 22137