谷歌研究团队于2026年11月13日发布了基于 Gemini 4 Argon、Carbon 和 Hydrogen 模型及 Antigravity /teamwork 多智能体工作流的最新科研成果 1。该系统通过结合并行探索与专用验证器(如 Lean 4),解决了数学、计算机科学、化学及软件工程领域的多个长期开放难题,包括证明信息论猜想、形式化韦尔猜想以及实现高性能内核优化 1。目前,相关工具已向外部研究人员开放,并提供免费访问通道以推动人机协作科研的发展 1。
在基础科学与理论计算机科学方面,该系统取得了多项突破性成果 1。它全量证明了信息论中的 Courtade–Kumar 猜想,并生成了5000万行 Lean 4代码 1。此外,系统自主完成了韦尔猜想(Weil conjectures)的端到端机器检查形式化,涉及约180万行 Lean 代码 1。针对量子计算领域,该系统还解决了一个关于 n-bit Boolean 函数查询复杂度的2014年开放问题 1。
在应用科学与工程优化层面,Antigravity /teamwork 同样展现了强大能力 1。在确定 HIV 蛋白酶抑制剂分子结构的任务中,系统达到了95%的 top-1 精确匹配率和100%的 top-3 准确率 1。软件工程方面,系统将 Zircon 微内核剩余 C++ 代码重写为 Rust,新增约80万行通过测试的代码,并使 libgav1 AV1 解码器移植后的速度提升了20%-30% 1。同时,训练出的9x9 Go 引擎在不到8天内 Elo 分数增加约2600,最终达到1885分 1。
为了促进学术界的人机协作研究,谷歌向10,000名学术研究人员提供了为期6个月的最高用量层级免费访问权限 1。
Google Research has announced the release of its latest scientific findings powered by the Gemini 4 model family—specifically Argon, Carbon, and Hydrogen—and the Antigravity /teamwork multi-agent system 1. Published on November 13, 2026, this initiative leverages parallel exploration combined with specialized verifiers such as Lean 4 to address persistent open problems across mathematics, computer science, chemistry, and software engineering 1. The company reports that these tools are now accessible to external researchers, marking a significant step toward human-machine collaborative scientific discovery 1.
The system achieved several major breakthroughs in formal verification and algorithmic optimization. In the realm of information theory, the team provided a full proof of the Courtade–Kumar conjecture, generating 50 million lines of Lean 4 code 1. Additionally, the AI autonomously completed the end-to-end machine-checked formalization of the Weil conjectures, involving approximately 1.8 million lines of Lean code 1. In quantum computing, the system resolved a 2014 open problem concerning the query complexity of n-bit Boolean functions 1. These achievements demonstrate the capability of the /teamwork workflow to handle complex logical structures and long-standing theoretical questions 1.
Beyond pure mathematics, the technology delivered significant results in applied sciences and software engineering. For chemical structure determination, specifically regarding HIV protease inhibitor molecules, the system attained a 95% top-1 exact match rate and a 100% top-3 accuracy rate 1. In software optimization, the team rewrote the remaining C++ code of the Zircon microkernel into Rust, adding roughly 800,000 lines of test-passed code 1. Furthermore, the porting of the libgav1 AV1 decoder resulted in a speed improvement of 20–30% 1. The system also trained a 9x9 Go engine, which saw its Elo score increase by approximately 2,600 points to reach 1,885 in less than eight days 1. To support broader research efforts, Google is offering six months of free access at the highest usage tier to 10,000 academic researchers 1.
评论
还没有评论,欢迎留下第一条。