xVerify: Efficient Answer Verifier for Reasoning Model Evaluations

Ding Chen, Qingchen Yu, Pengyuan Wang, Wentao Zhang, Zhengren Wang, Bo Tang, Feiyu Xiong, Xinchi Li, Chao Wang, mingchuan yang, Zhiyu li

Keywords: llm evaluation, reasoning model, slow thinking, judge model, reward model

TL;DR: We propose xVerify, a high-accuracy verifier for reasoning models that addresses the limitations of existing evaluation models and reward models in assessing the correctness of long CoT responses, outperforming all existing evaluation methods.

Abstract: With the release of the o1 model by OpenAI, reasoning models adopting slow thinking strategies have gradually emerged. As the responses generated by such models often include complex reasoning, intermediate steps, and self-reflection, existing evaluation methods and reward models are often inadequate. They struggle to determine whether the LLM output is truly equivalent to the reference answer, and also have difficulty identifying and extracting the final answer from long, complex responses. To address this issue, we propose xVerify, an efficient answer verifier for reasoning model evaluations. xVerify demonstrates strong capability in equivalence judgment, enabling it to effectively determine whether the answers produced by reasoning models are equivalent to reference answers across various types of objective questions. To train and evaluate xVerify, we construct the VAR dataset by collecting question-answer pairs generated by multiple LLMs across various datasets, leveraging multiple reasoning models and challenging evaluation sets designed specifically for reasoning model assessment. A multi-round annotation process is employed to ensure label accuracy. Based on the VAR dataset, we train multiple xVerify models of different scales. In evaluation experiments conducted on both the test set and generalization set, all xVerify models achieve overall F1 scores and accuracy exceeding 95%. Notably, the smallest variant, xVerify-0.5B-I, outperforms all evaluation methods except GPT-4o, while xVerify-3B-Ib surpasses GPT-4o in overall performance. Furthermore, we conduct RL experiments with xVerify as the reward model. Compared with direct generation, it shows an improvement of 22.9% for Qwen2.5-7B. which is greater than when Math Verify is used as the reward function. These results validate the effectiveness and generalizability of xVerify. All resources for xVerify are available at GitHub.

Primary Area: datasets and benchmarks

Submission Number: 11680