Title: SCI-PRM: A Tool-Aware Process Reward Model for Scientific Reasoning Verification

Abstract:

Although Process Reward Models (PRMs) have shown significant promise in the realm of mathematical reasoning, their potential within intricate scientific fields—such as physics, chemistry, and biology—has yet to be fully realized. Addressing scientific challenges requires more than just logical precision; it also demands strict factual accuracy and the correct application of specialized domain tools. Unfortunately, existing models frequently struggle with these areas, often resulting in hallucinations and a lack of robust verification. To bridge this gap, we introduce SCIPRM70K, a comprehensive new dataset that utilizes Chain-of-Tool trajectories to explicitly alternate between reasoning steps and the execution of scientific tools. Leveraging this dataset, we developed Sci-PRM, an efficient reward model designed to offer detailed, step-by-step supervision during a single inference pass. This supervision focuses on the accuracy of tool selection, execution, and result interpretation. Our experimental results indicate that Sci-PRM substantially improves foundation models in two primary ways: first, it facilitates effective test-time scaling through Best-of-N selection; and second, when employed within Reinforcement Learning frameworks, it provides a dense reward signal. This dense signal effectively counteracts the prevalent issue of advantage disappearance, empowering models to surpass current performance limits.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC