研究人员开发了 SciCrafter，一个在 Minecraft 中用于测试 AI 代理弥合科学发现与实际应用之间鸿沟能力的新基准。该基准使用参数化红石电路任务，要求代理发现并应用因果规则来实现特定的照明模式。对 GPT-5.2、Gemini-3-Pro 和 Claude-Opus-4.5 等领先模型的评估显示，它们的成功率在 26% 左右停滞不前，这凸显了在识别知识差距方面的局限性，而不仅仅是应用现有知识。 AI
影响 识别出 AI 代理开发中的一个新瓶颈，将重点从解决问题转移到制定问题。
排序理由 介绍 AI 代理能力新基准的新学术论文。
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
arXiv:2604.24697v1 Announce Type: new Abstract: Discovering causal regularities and applying them to build functional systems--the discovery-to-application loop--is a hallmark of general intelligence, yet evaluating this capacity has been hindered by the vast complexity gap betwe…
Discovering causal regularities and applying them to build functional systems--the discovery-to-application loop--is a hallmark of general intelligence, yet evaluating this capacity has been hindered by the vast complexity gap between scientific discovery and real-world engineeri…
Discovering causal regularities and applying them to build functional systems--the discovery-to-application loop--is a hallmark of general intelligence, yet evaluating this capacity has been hindered by the vast complexity gap between scientific discovery and real-world engineeri…