ARC-AGI-3 接近被攻克：编码工具链让 Opus 5 拿下 96.2%

精选理由

Jeremy Berman 用 Claude Code 加 Opus 5 把 ARC-AGI-3 刷到 96.2%，没做什么 ARC 定制，思路很通用。

AI 摘要

Jeremy Berman 使用 Claude Code + Opus 5（high）在 ARC-AGI-3 上拿到 96.2% 准确率，pass@2 为 99.3%。方案仅包含一个动作命令和文件系统日志，几乎没有 ARC 特定设计。作者此前预测编码工具链可泛化 LLM，这次结果支持了该判断。

原文 · Amjad Masad

ARC-AGI-3 is nearly solved by merely adding a coding harness. As predicted, coding generalizes LLMs.

ARC-AGI-3 is nearly solved by merely adding a coding harness. As predicted, coding generalizes LLMs. Jeremy Berman @jeremyberman I got 96.2% on ARC-AGI-3 with Opus 5, and 99.3% pass @2 . The program is basically Claude Code + Opus 5 (high), one action command, and filesystem logs. Almost nothing ARC specific. github.com/jerber/arc-code 🔗 View Quoted Tweet 💬 0 🔄 0 ❤️ 4 👀 1135 📊 1 ⚡

查看原推