Salvato in:
Dettagli Bibliografici
Autori principali: Guo, Jiaxing, Yang, Wenjie, Zhang, Shengzhong, Xu, Tongshan, Du, Lun, Zheng, Da, Huang, Zengfeng
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:https://arxiv.org/abs/2506.06877
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866908418330066944
author Guo, Jiaxing
Yang, Wenjie
Zhang, Shengzhong
Xu, Tongshan
Du, Lun
Zheng, Da
Huang, Zengfeng
author_facet Guo, Jiaxing
Yang, Wenjie
Zhang, Shengzhong
Xu, Tongshan
Du, Lun
Zheng, Da
Huang, Zengfeng
contents Outcome-rewarded Large Language Models (LLMs) have demonstrated remarkable success in mathematical problem-solving. However, this success often masks a critical issue: models frequently achieve correct answers through fundamentally unsound reasoning processes, a phenomenon indicative of reward hacking. We introduce MathOlympiadEval, a new dataset with fine-grained annotations, which reveals a significant gap between LLMs' answer correctness and their low process correctness. Existing automated methods like LLM-as-a-judge struggle to reliably detect these reasoning flaws. To address this, we propose ParaStepVerifier, a novel methodology for meticulous, step-by-step verification of mathematical solutions. ParaStepVerifier identifies incorrect reasoning steps. Empirical results demonstrate that ParaStepVerifier substantially improves the accuracy of identifying flawed solutions compared to baselines, especially for complex, multi-step problems. This offers a more robust path towards evaluating and training LLMs with genuine mathematical reasoning.
format Preprint
id arxiv_https___arxiv_org_abs_2506_06877
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Right Is Not Enough: The Pitfalls of Outcome Supervision in Training LLMs for Math Reasoning
Guo, Jiaxing
Yang, Wenjie
Zhang, Shengzhong
Xu, Tongshan
Du, Lun
Zheng, Da
Huang, Zengfeng
Computation and Language
Outcome-rewarded Large Language Models (LLMs) have demonstrated remarkable success in mathematical problem-solving. However, this success often masks a critical issue: models frequently achieve correct answers through fundamentally unsound reasoning processes, a phenomenon indicative of reward hacking. We introduce MathOlympiadEval, a new dataset with fine-grained annotations, which reveals a significant gap between LLMs' answer correctness and their low process correctness. Existing automated methods like LLM-as-a-judge struggle to reliably detect these reasoning flaws. To address this, we propose ParaStepVerifier, a novel methodology for meticulous, step-by-step verification of mathematical solutions. ParaStepVerifier identifies incorrect reasoning steps. Empirical results demonstrate that ParaStepVerifier substantially improves the accuracy of identifying flawed solutions compared to baselines, especially for complex, multi-step problems. This offers a more robust path towards evaluating and training LLMs with genuine mathematical reasoning.
title Right Is Not Enough: The Pitfalls of Outcome Supervision in Training LLMs for Math Reasoning
topic Computation and Language
url https://arxiv.org/abs/2506.06877