Papers

AGPV:分步验证错误定位的数学自我纠错框架

Agent-Guided Process Verification for Mathematical Self-Correction

Why we picked it

一篇很硬的数学自纠错论文,思路是把解答拆成逐条命题来验证,找到第一个错就停下重推,比只查最终答案的方法多了完整性检查。在 DeepSeek、Qwen、GPT 四个模型上平均准确率 63.8%,比最强基线高 13.5 个点。做推理 agent 的开发者值得仔细读。

arXiv 论文提出 Agent-Guided Process Verification(AGPV),针对多步数学推理中一步出错沿错误路径传播、解法遗漏必要分支的问题。AGPV 将当前解法分解为有序的关键命题序列,逐条验证并在需要时调用共享数学工具箱,检测到第一个错误即停止;随后由总结 agent 将失败证据转化为诊断,引导纠正 agent 从该点重新推导,并对修订后的解法进入下一轮验证,同时检查解法完整性。

Body

论文还给出理论保证:当关键命题覆盖充分且局部验证可靠时,AGPV 能保证解法的正确性与完整性。实现上采用压缩长工作上下文、在单个问题内复用成功工具结果的方式控制开销。在五个数学推理数据集上,配合 DeepSeek、Qwen、GPT 四个模型评测,AGPV 平均准确率达 63.8%,比最强代表性基线的 50.3% 高出 13.5 个百分点,修复覆盖率也高于程序化验证类基线。

相比只检查候选答案是否满足题设的结果验证方法,AGPV 的过程级定位能发现被结果验证遗漏的必要分支缺失问题,对做推理模型和数学 agent 的开发者有直接参考价值。

Original · arXiv cs.AI

Agent-Guided Process Verification for Mathematical Self-Correction

Large language models have made substantial progress in mathematical reasoning, yet in complex multi-step reasoning, an error at any step may propagate along an incorrect path, and the resulting solution may omit necessary branches. Existing natural-language reflection methods often struggle to locate critical errors precisely, while outcome verification typically checks only whether the available candidate answers satisfy the problem conditions and can therefore overlook necessary branches. To address these limitations, we propose Agent-Guided Process Verification (AGPV) for mathematical self-correction. AGPV decomposes the current solution into an ordered sequence of key claims, verifies them individually while invoking a shared mathematical toolbox when needed, and stops upon detecting the first error. A summarization agent converts the failure evidence into a diagnosis that guides a correction agent to re-derive the subsequent solution from that point; AGPV then adapts the next verification iteration to the revised solution. AGPV also verifies solution completeness. We further prove that AGPV guarantees solution correctness and completeness when key claims provide sufficient coverage and local verification is reliable, provided that all key claims pass verification. The implementation compresses long working contexts and reuses successful tool results within each problem. We evaluate AGPV on five challenging mathematical reasoning datasets with four models from DeepSeek, Qwen, and GPT. AGPV achieves an average accuracy of 63.8%, outperforming the strongest among the representative baselines evaluated (50.3%) by 13.5 percentage points, with higher repair coverage than the programmatic-verification baselines.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include AGPV, DeepSeek, Qwen, GPT.

Source attribution: curated from arXiv cs.AI (@Rui Zou, Mengqi Wei, Xin Zhao, Ji-Rong Wen).