Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

GPT-6.2 Astra Tops Autoform-Hard With a New Lean Proof

AI summary

Vals AI’s latest Autoform-Hard leaderboard ranks GPT-6.2 Astra first, with 49 of 56 advanced formal-mathematics problems solved and an 87.50% accuracy rate. Claude Fable 5.2 and GPT-6.2 Sol each solved 48 problems and tied for second at 85.71%. According to Vals AI, Astra completed a Lean formal proof of the Smoktunowicz gap theorem without human intervention in 94 minutes, producing 1,794 lines of Lean code and introducing a new “capacity” argument rather than using the original proof’s three theorems.

In-depth insight

The published run logs and proof link make the benchmark result independently inspectable, while Vals AI’s review found no matching argument in the cited 1995 and 2006 papers.

刚刚,第三方评测机构Vals AI放出了高难数学基准Autoform-Hard的最新榜单。

56道博士与科研级数学难题,此前10个前沿大模型去测,结果一道都没做出来。

而在这轮测试里,Claude Fable 5.2和GPT-6.2 Sol双双做对48道,打成平手。

唯独GPT-6.2 Astra多做对了第49道题,拿下了榜单第一。

这道唯一拉开差距的第49题,正是2006年发在数学四大顶刊上的Smoktunowicz间隙定理。

在全程零人工干预的环境里,Astra只用了94分钟、写下1,794行Lean代码,就拿下了它的全球首个形式化证明。

#GPT-6.2 Astra#GPT-6.2 Sol#Claude Fable 5.2#Autoform-Hard#AutoformBench#Lean#Mathlib#Meta FAIR#Vals AI

Original · 7 min readRead the original
Group chat

No comments yet. Start the conversation.