Beyond the First Error: Process Reward Models for Reflective Mathematical Reasoning
This addresses a specific bottleneck in training PRMs for mathematical reasoning, offering incremental improvements in data efficiency and performance over current methods.
The paper tackles the problem of data annotation for Process Reward Models (PRMs) in long chain-of-thought reasoning, where existing methods incorrectly assume all steps after the first error are wrong, by proposing a novel annotation method that accounts for self-correction and reflection, resulting in a 7B PRM trained on 1.7 million samples that outperforms existing PRMs in metrics like search guidance, BoN, and F1 scores.
Many studies focus on data annotation techniques for training effective PRMs. However, current methods encounter a significant issue when applied to long CoT reasoning processes: they tend to focus solely on the first incorrect step and all preceding steps, assuming that all subsequent steps are incorrect. These methods overlook the unique self-correction and reflection mechanisms inherent in long CoT, where correct reasoning steps may still occur after initial reasoning mistakes. To address this issue, we propose a novel data annotation method for PRMs specifically designed to score the long CoT reasoning process. Given that under the reflection pattern, correct and incorrect steps often alternate, we introduce the concepts of Error Propagation and Error Cessation, enhancing PRMs' ability to identify both effective self-correction behaviors and reasoning based on erroneous steps. Leveraging an LLM-based judger for annotation, we collect 1.7 million data samples to train a 7B PRM and evaluate it at both solution and step levels. Experimental results demonstrate that compared to existing open-source PRMs and PRMs trained on open-source datasets, our PRM achieves superior performance across various metrics, including search guidance, BoN, and F1 scores. Compared to widely used MC-based annotation methods, our annotation approach not only achieves higher data efficiency but also delivers superior performance. Detailed analysis is also conducted to demonstrate the stability and generalizability of our method.