按年被提到
20252025 · 11 次2026
数学、逻辑与理论›AI 数学与定理证明·arXiv:2507.15855 · cs.AI

Winning Gold at IMO 2025 with a Model-Agnostic Verification-and-Refinement Pipeline

用与模型无关的验证-精炼流水线拿下 IMO 2025 金牌
Yichen Huang, Lin F. Yang · 2025

靠验证与精炼提示流水线,让通用大模型解出 IMO 2025 五题

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 12 人首次 2025-07-22 · 最近 2026-01-02
12不同的人
6活跃月份
12讨论串
19被引(OpenAlex)

作者设计了一套与模型无关的「验证—精炼」提示流水线,并在 IMO 2025 上评测,对发布早于比赛的模型避免了数据污染。配上 Gemini 2.5 Pro、Grok-4 或 GPT-5 中任意一个,流水线都能正确解出 6 题中的 5 题(约 85.7%),而基线在 32 个候选中取最优仅为 31.6%、21.4% 和 38.1%。结论是推理能力的提升不只靠更强的基座模型,也靠善用模型的方法。

HN 为什么提它

评论者用它说明:只靠提示与编排,公开可用的通用模型(如 Gemini 2.5 Pro)不借助外部工具也能解出大部分 IMO 2025 题目,常用来反驳对 LLM 数学能力的质疑。

实证研究难度 2 / 5约 20 分钟有时效性
读之前最好知道
1大语言模型提示工程
2LLM 数学推理与竞赛题(IMO)
3自我验证与迭代精炼

HN 书评

推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及
推荐jsnell · 2025-09

“have a look at this paper where some floks used the normal generally available Gemini Pro 2.5 to solve 5/6 of the 2025 IMO problems”

推荐它,用普通 Gemini Pro 2.5 解出了 IMO 2025 六题中的五题
DeepMind and OpenAI win gold at ICPC原评论 ↗
引用bwfan123 · 2025-07

“This paper [1] shows that gemini pro 2.5 without data-contamination and some minimal prompting (llm orchestration) can solve problems 1-5 on imo 2025.”

拿它说明无数据污染、简单提示编排下 Gemini 2.5 Pro 能解出 IMO 1-5 题
Gemini with Deep Think achieves gold-medal standard at the IMO原评论 ↗
引用demirbey05 · 2025-08

“IMO is not breakthrough, if you craft proper prompts you can excel imo with 2.5 Pro. Paper: Winning Gold at IMO 2025 with a Model-Agnostic Verification-and-Refinement Pipeline.”

拿它说明 IMO 不算突破,提示词写得好 2.5 Pro 也能做出来
引用986aignan · 2025-08

“[2] Winning Gold at IMO 2025 with a Model-Agnostic Verification-and-Refinement Pipeline Appendix A”

指引去看论文附录 A 里 Gemini 的 IMO 2025 证明
Let's properly analyze an AI article for once原评论 ↗
引用starchild3001 · 2025-08

“Reference: Winning Gold at IMO 2025 with a Model-Agnostic Verification-and-Refinement Pipeline”

贴出它作为参考出处
Claim: GPT-5-pro can prove new interesting mathematics原评论 ↗
引用yorwba · 2025-10

“I think you meant to link to”

指出对方想贴的应是另一个链接
Implicit Actor Critic Coupling via a Supervised Learning Framework for RLVR原评论 ↗
看全部 9 条 →
引用impossiblefork · 2025-10

“It's reasonable to question it, but there's a fun Chinese paper where they attempt to create a system of prompts that can be used with commercial LLMs to solve hard maths problems.”

推荐这篇有趣的中国论文,用一套提示让商用 LLM 解难数学题
It's OpenAI's world, we're just living in it原评论 ↗
引用NitpickLawyer · 2025-11

“a) no, gemini 2.5 was shown to "win" gold w/o tools. - Winning Gold at IMO 2025 with a Model-Agnostic Verification-and-Refinement Pipeline”

反驳并拿它说明 Gemini 2.5 不用工具也能拿下金牌水平
Yann LeCun to depart Meta and launch AI startup focused on 'world models'原评论 ↗
引用CamperBob2 · 2026-01

“How can anyone choose to remain so ignorant in the face of irrefutable evidence that they're wrong?”

拿它当无可辩驳的证据,反问怎能对此视而不见
2025: The Year in LLMs原评论 ↗
被提交为 story · 3 次