Process Reward Models (PRMs) have become essential for enhancing the reasoning capabilities of Large Language Models (LLMs) by providing fine-grained feedback at each intermediate step. Recently, generative verifiers have emerged as a promising direct...
Process Reward Models (PRMs) have become essential for enhancing the reasoning capabilities of Large Language Models (LLMs) by providing fine-grained feedback at each intermediate step. Recently, generative verifiers have emerged as a promising direction, leveraging the inherent reasoning power of LLMs to verify solutions via Chain-of-Thought (CoT). However, existing generative PRMs suffer from overconfidence, often assigning extreme probability scores to reasoning steps, which undermines their reliability in downstream applications. In this thesis, we introduce the Uncertainty-Aware Process Reward Model (UAPRM), a novel framework designed to mitigate this issue by explicitly modeling confidence within the verification process. We propose a methodology to train generative verifiers that output both a verification rationale and a calibrated confidence score, utilizing a self-consistency-based data augmentation pipeline. Furthermore, to address the inherent data imbalance where high-confidence samples dominate, we employ a token-wise weighted loss strategy that forces the model to learn representations of uncertainty more effectively. Our experiments on mathematical reasoning benchmarks demonstrate that UAPRM outperforms baseline generative verifiers in terms of verification precision and calibration error. We show that these calibrated confidence scores translate into superior performance in Best-of-N selection tasks, establishing UAPRM as a more reliable and robust verifier for complex reasoning.