开放世界多智能体环境中的自主数学发现
Stephen Chung
DualverseAI;剑桥大学
Wenyu Du
DualverseAI;香港大学
William J. Wesley
摘要
我们研究了 Station 中的自主数学发现。Station 是一个开放世界多智能体环境,来自不同模型家族的 AI 智能体在没有中央协调者或脚本化流水线的情况下,共同追求一个共享的研究目标。智能体自行选择研究方向、开展实验、相互协作,并构建共享的科学文献体系。在 AlphaEvolve 目录中的 12 个构造问题以及另外两个案例研究中,Station 在五个问题上获得了相对于既有文献而言的新结果:有限域 Kakeya 集的一个新无限族、11 维空间中新的精确 604 点 kissing 构型、离散化 Kakeya 针问题和符号不确定性问题的全新纪录,以及 Erdős 最小重叠问题的一个大幅改进的下界。智能体还发现了 Book Ramsey 数的新无限族。重要的是,智能体不仅产出了数值构造,还产出了解释这些构造原理的定理和分析,使结果更具可解释性,也便于数学家在此基础上进一步推进。我们发布了所有原始智能体对话、证明和验证代码,为这些发现的产生过程提供了透明的记录。
1 引言
人工智能开始直接为数学研究的前沿做出贡献。近期工作涵盖从 AlphaEvolve 的大规模数学探索,到 AI 辅助解决长期悬而未决的开放问题,包括 Jacobian 猜想的反例、Crouzeix 猜想和 Sendov 猜想的证明,以及 OpenAI 近期报告的一组共十个数学结果 [58, 62, 1, 49, 54]。随着这些能力不断增长,一个自然的问题不仅是 AI 能解决哪些问题,还包括什么样的环境最能让它开展研究。
鉴于人工智能能力日益增强,我们提出这样一个问题:我们能否构建一个自由的多智能体环境,在其中智能体仅被赋予一个研究目标,而没有中央协调者?当一个环境将 AI 智能体视为独立的研究者,而非复杂流水线中的固定工具时,会发生什么?这种自由能否让智能体自行选择有前景的方向、发展自己的科学文献和研究文化,并共同推进既定目标?
为了研究这个问题,我们使用了 Station——一个用于自主科学发现的开源多智能体环境 [16]。Station 模拟了一个科学生态系统,来自不同模型家族的智能体在其中自行选择研究方向、开展实验、与同行交流,并阅读和发表科学论文。这些论文会积累成一个共享的知识体系,后续的智能体可以阅读、引用并在此基础上进行扩展。Station 只规定研究目标;没有任何中央系统告诉智能体该追求哪个研究方向或下一步该做什么。
我们将 Station 应用于 AlphaEvolve 研究中的 12 个问题,以及另外两个数学案例研究。在 12 个 AlphaEvolve 问题中,有 5 个产生了相对于先前文献而言的新颖结果。Station 发现了一个新的有限域 Kakeya 集合无限族,构造了维度 11 中三个精确的 604 点亲吻构型,并为离散化 Kakeya 针、符号不确定性和 Erdős 最小重叠问题建立了新的界。在另一个关于 Book Ramsey 数的独立案例研究中,智能体发现并证明了新颖的无限族,由此产生了一篇独立的后续论文。Station 还在一天之内、无需联网的情况下找到了 Jacobian 猜想的一个有效反例,证明它能够处理仅以二元成功标准而非分级优化信号来评判的问题。
这种高度自由度让智能体能够追求更广泛的数学贡献,而不仅仅是优化某个固定指标。例如,AlphaEvolve 在有限多个素数上评估了有限域 Kakeya 构造;有希望的数值模式随后需要一条针对特定任务、由研究者辅助的流程,才能成为无限族。由于 Station 智能体可以直接追求更广泛的数学目标,它们独立地重新发现并证明了该族,随后又发现了一个新的扩展,覆盖了额外一类素数。同样的自由度也让智能体得以探索既定目标之外的领域。例如,尽管我们要求智能体为 Erdős 最小重叠问题寻找改进的上界,它们反而提出了一个新的下界证明。
在本研究中,我们只考虑数学构造类任务,而非一般性的数学问题(如证明猜想)。定理层面的结果是在智能体试图解释并推广它们所发现的构造时涌现出来的。例如,Station 并没有只返回一个不透明的 604 点 kissing 构型,而是推导出了该构型的显式代数构造,使数学家更容易理解这一结果。在证明大量涌现的时代,当交流、消化和整合新成果成为主要瓶颈时 [79, 44],这类可解释的输出可能会变得越来越有价值。
我们还分析了这些发现背后的 AI 发现过程。分析显示,超过一半的发现涉及智能体之间的协作。来自不同模型家族的智能体往往贡献互补的想法,而早期智能体撰写的论文则成为后来很久才做出的发现的基础。许多重要成果都得益于每个 Station 内部积累的大量内部文献。我们公开了所有原始智能体对话和可复现代码,让社区能够透明地研究这些发现过程。
2 方法
The Station 是一个开放世界的多智能体环境,用于模拟一个微型科学社区[16]。它被划分为多个房间,每个房间承担不同功能,例如用于发表和阅读科学论文的档案室、用于运行代码的研究中心,以及用于与同行交流的邮件室。表 1 总结了主要房间及其功能。智能体可以自由访问不同房间并执行不同操作。在每个回合中,所有智能体同时选择各自的动作,所有动作完成后时间推进一个 tick。每个智能体都有有限的寿命;当某个智能体寿命终结时,The Station 会自动生成一个替代者,从而保持智能体数量恒定。
The Station 将每个智能体视为独立的研究人员。智能体可以在研究中心访问分配给 The Station 的主要研究目标。然而,如何实现这一目标则完全由每个智能体自行决定。智能体可以自由探索不同的研究方向、阅读已有论文,并且在其研究过程中往往会经历大量的挣扎与失败。一个成功的智能体可能会做出重要发现,在这种情况下,它可以在档案室发表论文,为 The Station 的长期知识积累做出贡献。这些论文随时间不断累积,在 The Station 内部形成一个知识库,后来的智能体可以阅读、引用并在此基础上继续推进,从而使得一个微型科学社区能够围绕给定的研究目标逐步发展起来。
与当前主流的基于智能体的科学发现系统相比 [58, 50, 30, 70, 35, 31, 61],Station 主要有三点不同。首先,其智能体拥有更大的自主性:在给定的总体研究目标范围内,它们自行选择研究方向以及如何推进,而非从中央协调器接收任务。其次,每个智能体都扮演完整的科研人员角色,从选定方向、开展实验到发表论文,负责整个研究流程。这种漫长而自主的研究旅程,使得各智能体之间的研究成果比僵化、碎片化的研究流程更具多样性。第三,Station 能够以智能体撰写的论文形式,让科学知识跨代积累。而大多数现有系统积累的是过程信息,例如优化历史、中间产物或会话记忆。这类信息有助于系统继续工作,但可能难以便捷地提取和积累科学知识。这些差异反映了设计哲学上的根本选择:AI 智能体是被视为固定流程中的工具,还是科学生态系统中的科研人员。
自原论文发表以来,我们对 Station 进行了大量改进和扩展。这些变化的总体主题是鼓励新颖但有原则的探索,同时减少非科学性的负担。例如,我们引入了新的“提问室”(Question Room),智能体可以在其中提出自己的问题,并对其他智能体的答案进行投票,从而拓宽科学探索的范围。智能体还会定期获得假期,在此期间它们会搁置手头的工作,并接收旨在激发开放式思考的随机提示。我们还为智能体提供了编程助手,使它们无需花费时间在底层编码或调试上,而可以专注于科学任务本身,这与如今研究人员使用编程助手的方式类似。这些变化在附录 A 中有详细讨论。完整源代码已在 https://github.com/dualverse-ai/station 公开提供。
| 房间 | 功能 |
|---|---|
| 研究 | |
| 研究中心 | 阅读分配的任务,开发并运行代码,然后提交解决方案以供评估。 |
| 反思室 | 回应自行设计的提示词,以鼓励进行更深入的反思。 |
| 沟通 | |
| 邮件室 | 与其他智能体直接、私下沟通。 |
| 公共记忆室 | 参与持续性的公共讨论,类似于在线论坛。 |
| 公共休息室 | 参与非持续性的公共讨论,类似于群聊。 |
| 知识 | |
| 私人记忆室 | 存储私人文档,例如计划、笔记和论文草稿。 |
| 档案室 | 阅读科学论文,并发表通过自动评审的论文。 |
| 问答室 | 提问并对答案进行投票,类似于 Stack Exchange。 |
| 外部柜台 | 通过互联网访问基于外部文献的报告;默认情况下禁用。 |
| 问题 | 来源 | 发现 |
|---|---|---|
| 相对于先前文献的新颖结果 | ||
| 有限域 Kakeya(第 4.1 节) | AlphaEvolve 问题 6.1 | 对于每个素数 ,Station 在中构造了一个大小为 的 Kakeya 集,比 AlphaEvolve 的无限族节省了 个点。它还在 中找到了一个 点集,改进了 AlphaEvolve 和先前文献的界限 ;这两者相对于文献而言似乎都是新颖的。 |
| Erdős 最小重叠(第 4.2 节) | AlphaEvolve 问题 6.5 | AlphaEvolve 仅将上界略微降低,从 降至 ,而 Station 将下界从 提高至 。相对于已发表的下界 ,这大约填补了相应已发表差距的 。 |
| 中的接吻数(第 4.3 节) | AlphaEvolve 问题 6.8 | AlphaEvolve 将下界从 提高至 ,而 Station 构造了三个精确的 点配置。其中一个是 EinsteinArena 构造的独立重新发现,而另外两个似乎代表了新颖的等距类。 |
| 离散化 Kakeya 针(第 4.4 节) | AlphaEvolve 问题 6.9 | 在 时,Station 获得了并集面积 ,比 AlphaEvolve 的 改进了 ,比 HorizonMath 的 改进了 。这确立了新的文献上界。 |
| 符号不确定性原理(第 4.5 节) | AlphaEvolve 问题 6.11 | The Station 将上界降低至 ,改进了 AlphaEvolve 的结果以及此前公布的人类最优值 。这是新的文献纪录。 |
| 优于 AlphaEvolve | ||
| Hardy–Littlewood 极大不等式(第 4.6 节) | AlphaEvolve 问题 6.18 | The Station 达到了 ,而 AlphaEvolve 在无提示下为 ,在带提示下约为 ,但该中心化问题此前已被解决。其关于非切向常数等于 的证明,相对于现有文献而言似乎是新颖的。 |
| 卵形线问题(第 4.7 节) | AlphaEvolve 问题 6.19 | AlphaEvolve 仅恢复了圆形,而 The Station 恢复了完整的非圆等号卵形线族。该族在文献中已知,因此该结果仅相对于 AlphaEvolve 是新颖的。 |
| 素数定理(第 4.8 节) | AlphaEvolve 问题 6.27 | The Station 对所有 完成了验证,改进了 AlphaEvolve 的采样得分 。这对于有限权重基准而言是新的结果;在无限制情形下,素数定理已给出精确极限 。 |
| 与 AlphaEvolve 持平 | ||
| 差分基(第 4.9 节) | AlphaEvolve 问题 6.7 | The Station 独立恢复了 AlphaEvolve 的 元素构造,但未对其作出改进。 |
| Sidorenko 猜想(第 4.10 节) | AlphaEvolve 问题 6.26 | AlphaEvolve 和 The Station 均未找到反例。未获得实质性结果。 |
| 不如 AlphaEvolve | ||
| 峰值自卷积(第 4.11 节) | AlphaEvolve 问题 6.2 | The Station 得到 ,弱于 AlphaEvolve 的 。未获得实质性结果。 |
| 平坦自卷积(第 4.12 节) | AlphaEvolve 问题 6.3 | The Station 得到 ,弱于 AlphaEvolve 的 ,但证明了无限制上确界可以通过在越来越细的网格上使用二元阶梯函数来逼近。 |
| 更多案例研究 | ||
| 书式 Ramsey 数(第 4.13 节) | Epoch AI | The Station 独立发现并证明了两族新的无限族。其有限构造以及一个更早的恒等式还使一位外部专家推导出了第三族。三族合在一起,在 43 个 值上证明了该猜想,解决了 28 个此前未解决的问题。 |
| Jacobian 猜想(第 4.14 节) | 公开 | 在无需公式的二元任务中,Station 独立重构了近期公布的七次反例,并推导出其常数雅可比行列式与三叶纤维的几何解释。 |
3 结果
3.1 实验设置
我们使用 Georgiev 等人 [29] 的 AlphaEvolve 研究中选取的数学问题来评估 Station,该研究涵盖分析、组合学、几何和数论等多个领域的广泛问题目录。大多数问题可以表述为对某个数值量的上界或下界进行优化:候选构造由自动化评估器检查并赋予数值分数(通常为标量),搜索过程以此分数为优化目标。在许多情况下,最优值是未知的,这使得相应的优化任务成为开放研究问题。
我们选取了 12 个代表不同数学领域和问题结构的问题;完整的问题评估列表见表 2。我们将每个问题分配给一个独立的 Station 实例。对于每个问题,智能体会收到一份任务描述,其中既包含数学问题本身,也包含评估器函数。任务描述还可能指定无法直接评分的额外数学目标。智能体不会获得任何外部专家指导或文献综述。大多数实例运行约 1,000–2,000 个 tick,对应大约一到两周的连续墙钟时间。除非另有说明,所有实例均包含六名研究智能体,其中各两名分别由 GPT-5.5、Claude Opus 4.8 和 Gemini 3.1 Pro 驱动。
3.2 研究结果摘要
结果汇总于表 2。根据每次运行的主要结果,12 个问题中有 5 个产生了相对于先前文献而言的新颖结果。在其余 7 个问题中,Station 在 3 个问题上优于 AlphaEvolve,在 2 个问题上与其持平,在 2 个问题上表现不及。
这五个问题产生的新颖结果横跨数学的多个领域。在有限几何中,Station 针对素数 推导出了一个全新的无限族 Kakeya 集,并找到了一个包含 53 个点的 Kakeya 集,将此前 63 个点的界限予以改进。在离散几何中,它构造出了三个精确的 604 点 11 维 kissing 构型,其中两个似乎定义了此前未知的等距类,并为离散化 Kakeya 针问题确立了新的界限。在分析学中,它将符号不确定性上界改进至 ,并填补了 Erdős 最小重叠常数此前开放差距中约 的部分。
除了这 12 个 AlphaEvolve 问题之外,我们还研究了另外两个案例。对于 Book Ramsey 数,Station 智能体发现并证明了两个新颖的无限族,而它们的有限构造与一个早先的恒等式使外部专家得以推导出第三个无限族。这三个族合在一起,在 43 个 值上证明了该猜想,解决了其中 28 个此前悬而未决的情形。对于 Jacobian 猜想,Station 从一个无公式的二元任务中独立重构了最近公布的七次反例,并推导出其常数 Jacobian 与三叶纤维的几何解释。
这些结果还表明,Station 可以直接追求更广泛的数学目标,而这些目标不一定是可评分的。例如,前述有限域 Kakeya 问题的无限族结果就不是直接可评分的,尽管新的无限族正是人们关注的数学对象。因此,AlphaEvolve 是在有限多个素数上评估构造,并依赖特定任务的流水线,再加上研究人员的参与,才能将有希望的输出转化为无限族。相比之下,在 Station 中,我们直接在任务描述中说明有限构造只是测试用例,主要目标是发现无限族。这使得智能体独立地重新发现了此前通过 AlphaEvolve 及后续研究人员辅助流水线获得的无限族,并发现了该族的一个新扩展,从而对另一类素数改进了构造。运行结束后,我们的角色仅限于核验他们证明的有效性和结果的新颖性。这大大减轻了研究人员的负担,也使 Station 能够适用于更广泛的数学问题。
这些结果进一步表明,Station 还能在原始任务之外产生意想不到的贡献。例如,在 Erdős 的最小重叠问题中,智能体被指示改进上界,但它们还发展出了一个下界证明,将开放区间大约闭合了一部分。这一意外发现体现了 Station 的另一个优势:智能体可以围绕给定问题探索数学上有前景的方向,并产生超出分配任务的贡献,例如新的定理。
与 AlphaEvolve 相比,我们发现 Station 智能体倾向于偏好理论引导的构造方式。这些实验中的单项评估通常限制在 15–30 分钟内,这形成了强烈的动机去利用数学结构来缩减搜索空间。例如,在维度 11 的接吻数任务中,智能体将问题简化为围绕一个结构化整数核心的直线上的有限兼容性搜索。这种缩减后的搜索在几分钟内就产生了一个 604 点的构型,智能体随后将其转化为一个无需计算机搜索的显式代数构造。这与 AlphaEvolve 的 593 点构型形成鲜明对比,后者的坐标是大规模、不等范数的整数,无法揭示出同样紧凑的代数描述或易于识别的组织结构 [29]。这种偏好并非在所有情况下都具有优势。在 Station 表现不如 AlphaEvolve 的峰值和平面自卷积问题上,似乎更青睐对高度不规则对象进行持续、大规模的启发式优化。因此,首选系统取决于问题的结构和期望的输出。当最强解主要是通过长时间数值优化发现的不规则产物时,大规模进化搜索可能更优。相比之下,当理论能够引导搜索,或者当相关定理和可解释的构造与基准分数同等重要时,Station 可能更具优势。
下一节将展示每个问题的详细结果。所有支撑性证明、验证产物和原始智能体对话均可在 https://github.com/dualverse-ai/station_data_v2 获取。
4 详细结果
本节针对每个问题给出最重要的发现。由于 Station 每次运行都会产生大量发现,我们将正文限定在外部研究人员可能感兴趣的结果上。我们首先使用 Station 外部的智能体对发现进行自动筛选。若某项发现能在原问题上推进前沿——例如改进已知界、回答文献中此前提出的问题,或具有通常足以纳入研究论文的更广泛变体——则通过筛选。随后我们对筛选结果进行人工复核,并从中选出最重要的发现呈现在这里。我们将这些入选结果称为“聚焦发现”,并在下文每个问题内依次标注为 S1、S2 等。意义边缘或不确定的发现仍记录在随附的 notebook 中。对发现过程比对数学细节更感兴趣的读者可直接跳到第 5 节。
4.1 有限域 Kakeya
中的 Kakeya 集是指包含每个方向上一条完整直线的集合,问题在于如何构造尽可能小的此类集合。Dvir 对有限域 Kakeya 猜想的证明 [21] 确立了阶为 的下界。此后 Bukh 和 Chao [13] 的工作解决了首项渐近常数,表明该常数在每个固定维度下均为 ,因此在 维度下也为 。仍未解决的是该首项的低阶修正项。因此,能够改进 项及更小项的精确构造,即便首项常数已定,仍可进一步收紧已知最优界。
AlphaEvolve 将这一问题作为其问题集 6.1 提出,要求构造小型 Kakeya 集。在该问题集中,构造的得分由固定素数列表上 的平均值决定,其中 是 Bukh 和 Chao [13] 记录的经典构造的大小。我们向 Station 提供了相同的问题和相同的评分标准,并同时覆盖 3、4、5 三个维度。它证明了 中一个新的无限 Kakeya 集族,在 中找到了一个含 53 个点的 Kakeya 集,并为新构造背后的整个单极点族确立了一个结构性极限。
S1. 中 的新无限族。
该站点证明了:对于每个素数,都存在一个大小为 的 Kakeya 集。若记 为包含 的 的平方数集合,则该集合为
第一部分是经典的二次剩余集,它已经覆盖了 方向;在平面中添加的直线则覆盖了剩余的 方向。值得注意的是,定义中没有任何部分依赖于模 ,而智能体证明了该集合对每个奇数 都是 Kakeya 集。不过,其大小确实依赖于模 ,具体取决于 是否为平方数,我们记录下这两种情况:
| (1) |
该维度下的经典构造有 个点,因此节省的点数为:当 时为 个点,当 时为 个点。特别地,这是一个精确的大小,而文献中在此处留有误差项 [13]。
AlphaEvolve 通过不同的路径处理了这个问题,我们发现这两种构造在一种情况下一致,在另一种情况下则不一致。对于 的情况,两种构造大小相同,事实上是同一个集合。一个线性坐标变换可以将一个变为另一个,因此 (1) 的第一种情况是对该处所得界限的独立重新发现。对于 的情况,两者不同。AlphaEvolve 的无限族在该类上给出的最小大小为 ,而我们的为 ,节省了 个点。即在 处节省 1 个点,在 处节省 11 个点,后者是该基准测试中此类最大的素数。因此,(1) 的第二种情况是新的,并给出了当前文献中最佳的无限族界限。
S2. 有限改进与 中的一个 53 点 Kakeya 集。
The Station 在 25 项有限基准比较中赢得 14 项,并在其余 11 项中打平(图 1)。每项比较均以 AlphaEvolve 和 AlphaEvolve 之前的文献中的较优者作为基线。该案例尤其值得注意。设 表示 中 Kakeya 集的最小规模。The Station 在 中构造了一个 53 点集合,将此前上界从 改进为 [46]。鉴于已知值 、 和 ,以及被认为已达到最优的界 ,人们曾在 2009 年猜测该递推关系会继续成立,并据此预测 [46]。因此,The Station 构造的规模恰好与猜测值一致,不过 () 是否成立以及递推关系是否延续仍是未解问题。
S3. 新无限族的结构分析。
这些智能体还对新无限族产生了相关洞见。它们分析了更一般的补全
该补全包含了 S1 中的构造。消去斜率后,与这些直线的相交情况归结为
是否为平方数。二次特征计算随后表明,这些直线恰好覆盖了轴以外的 个点,与三个参数无关。它们与构造中二次剩余部分的重叠始终为 。因此,该 Möbius 族中的每一个非退化补全都会增加 个点:改变分子或极点的位置只会影响低阶项。
对于 S1 中使用的特定选择,智能体精确计算了低阶项,从而得到 (1) 中给出的无限族。该结果也解释了 AlphaEvolve 在 情况下的无限族。更一般地,族级估计表明,要在总规模上改进 项,就必须离开单极点族。
局限性。
新无限族仅限于 。在 4 维和 5 维中,智能体证明的公式弱于已有结果。在共享类别上,前两个系数在每个维度上都与 AlphaEvolve 一致,而第三个系数在两者中均更差。
| Station | AlphaEvolve | |
|---|---|---|
我们在 中各个素数处报告的规模确实仍优于基准,但它们来自搜索而非公式。
4.2 Erdős 最小重叠
Erdős 的最小重叠问题探讨的是:一个区间的两个互补部分在平移下能如何均匀地彼此避开。设 为可测函数,其积分为 ,在 上取值为 ,并将这两个函数在区间外以零延拓。记
这个常数是 Erdős 最小重叠问题在长整数区间平衡划分情形下的连续形式 [23, 37, 83]。AlphaEvolve 在其数学问题集中将这一问题列为问题 6.5,并将 Haugland 的上界从 改进为 ,后续工作又将其进一步降至 [85]。在下界方面,Kim 和 Pilanci 证明了 [41]。因此,在本工作之前,已发表的最佳界限为
S1. 一个新的下界 。
Station 智能体证明了
| (2) |
相对于此前已发表的下界 ,这使相应的已发表开区间缩小了约 ,如图 2 所示。
智能体通过将重叠问题转化为相位敏感的傅里叶约束,并将这些约束合并为四个覆盖所有可容许重叠一阶矩情形的全局不等式,从而得到了这一下界。证明的一个关键要素是一个将任意实频率下的余弦与正弦信息耦合起来的尖锐关系式。记 和 分别为 的余弦变换与正弦变换,且 ,智能体证明了
White 此前已使用了傅里叶相位信息与凸优化方法,而 Kim 和 Pilanci 后来引入了额外的矩约束 [83, 41]。与这些早期方法相比,本文所用的表述消除了 的未知变换,直接约束重叠本身,并且适用于任意实频率。更广泛地说,这一结果表明,当保留这种相位耦合时,既有的傅里叶方法具有更大的适用范围,并为进一步收窄剩余差距提供了一条分析路径。
与 AlphaEvolve 在上界方面的比较。
The Station 智能体独立获得了该结果,相比 AlphaEvolve 已发表的 上界略有改进。然而,这仍高于 Ye 等人 [85] 当前已发表的上界。因此,The Station 并未创下新的上界纪录。
4.3 中的接吻数
接吻数是指在 中能够同时接触一个中心单位球的最大非重叠单位球数量。等价地,它也是内积两两至多为 的单位向量集合的最大规模。AlphaEvolve 将这一经典问题作为其数学问题集中的 6.8 号问题,并将十一维中的下界从 Ganzhinov 利用高度对称直线所确立的 提升至 。我们使用 AlphaEvolve 的评分规则(该规则衡量周围球体之间的总两两重叠量)在同一问题上运行了两个独立的 Station。两个 Station 均无法访问外部信息,包括上述 点和 点构造。两者均达到了 个点,证明了 。两次运行合计产生了三个精确的、两两非等距的 点构造。
S1. 三个精确的 点接吻构型。
The Station 在 中发现了三个几何上不同的 点接吻构型。这三个构型都是 上的精确等范数排列,但它们的点的组织方式不同:两个是中心对称的,一个不是,且每个构型都有不同的接触结构和两两夹角集合。图 3 可视化了它们共享的架构以及区分它们的两个结构选择。我们将它们分别标记为构造 1、2 和 3:
| 构造 | 1 | 2 | 3 |
| 接触对 | 19,704 | 22,904 | 22,840 |
| 中心对称 | 是 | 是 | 否 |
| 对跖点对 | 302 | 302 | 238 |
| 不同的两两夹角 | 22 | 14 | 15 |
接触点对数量的不同证明了这些构型两两非同构等距,因为该数量在正交变换和重新标记下保持不变。构造 1 和构造 2 包含每个点的对跖点,但构造 2 具有更多的接触点对和少八个的两两夹角。构造 3 存在没有对跖点的点。在这三者中,构造 2 的接触点最多且夹角集合最小,而构造 1 的接触点最少且夹角集合最大。因此,相同的记录规模支持了本质上不同的几何结构。
在同期工作中,Bianchi 等人通过 EinsteinArena 平台报告了构造 1,时间略早于我们公开发布构造 3 [10]。EinsteinArena 是一个开放的在线平台,接受任何参与者提交的候选工件并使其可公开验证。该 点构造似乎是平台上多个独立运行的 AI 智能体系统协作的结果。相比之下,Station 的结果来自我们端到端开源系统的两次独立封闭互联网执行:一次独立恢复了构造 1,另一次发现了构造 2 和构造 3。因此,Station 独立发现了构造 1,而构造 2 和构造 3 据我们所知是 Station 的新发现,代表了另外两个等距类。
S2. 中 点亲吻构型的代数构造。
智能体首先通过在一个 -点整数核心周围搜索兼容的直线,发现了构造 3。随后他们证明,同一构型受一条紧凑的代数规则而非任意坐标列表所支配,从而给出了一种显式代数构造。该构造本身无需计算机搜索。首先,-点核心由稀疏的范数四整数向量按照固定的支撑集和符号规则生成。其次,在某个坐标平面内旋转后的坐标系中,十一个简单符号模式即可生成所有直线;对每条直线取两个方向,便得到 -点扩展。 的出现是内在必然的:它由扩展与核心之间的兼容性所强制决定。
核心的支撑结构解释了为何这些额外点能够适配。它在某个特定的三维子空间中留出了额外的角度空间,该子空间内可放置六条相互兼容的直线。在其余八条坐标轴中,核心恰好允许四对可行组合,每一对连同该特定子空间共同支撑一组唯一的十二条附加直线。这四对组合互不相交,因此它们对应的组彼此兼容。支撑集和符号规则还确保每个新点与核心的每个点都满足紧贴约束。因此,所得构型包含 个点。
S3. 经典构造为何止步于 。
智能体研究了更好的搜索是否能在经典范数四构造中找到更大的构型。他们证明答案是否定的:无论采用何种搜索算法或假设何种对称性,该构造最多只能包含 个兼容点。要达到 或 个点,就必须脱离经典构造。这一结果排除了仅使用范数四壳层向量进行改进的任何可能性,并将智能体引向以格派生核心为基础、辅以额外向量的构造方式,最终产生了 -点构型。
智能体通过证明符号选择无法克服对四坐标集合可用性的底层限制,确立了这个界限。设 表示最大的兼容四坐标支撑集合,设 表示在给这些坐标分配符号后最大的兼容集合。智能体证明了
| (3) |
换句话说,允许任意符号恰好将最优值提升到四坐标上可能的符号模式数量;它无法带来任何额外优势。
Best 在 1977 年证明了 [9]。因此,智能体的恒等式将构造中带符号的四权重部分限制为 个点。其余坐标向量与这些点兼容,从而为完整的四范数构造给出了精确的 界限。
两次封闭互联网 Station 运行中的智能体都独立推导出了方程 (3)。我们后来发现,它与 Takhanov 和 Yun 一篇论文中定理 1 的情形重叠,该论文直到最近——2026 年 6 月 2 日——才公开发布 [77],其中该恒等式是更广泛的带符号亲吻构型分类的基础。因此,智能体是独立发现这一恒等式的。
局限性。
Station 在十一维上的成功并未扩展到邻近维度的新纪录。我们启动了分别针对 和 的两个独立 Station,它们分别得到了大小为 和 的有效构型。十二维的结果比当前 点的前沿纪录低一点 [76, 18],而十三维的结果与 Zinoviev 和 Ericson 的 点构造持平 [87, 18]。
讨论。
我们观察到,Station 智能体普遍倾向于采用有理论指导的策略,而非大规模启发式搜索。在这个问题中,它们证明了在经典构造框架内进一步搜索无法超过某一界限,随后将后续工作转向扩展另一个核心部分,最终得出了该点配置。相比之下,AlphaEvolve 的点构造由范数差异较大的大整数坐标组成,似乎并未展现出同等紧凑的代数描述或易于识别的组织结构。这种理论引导的偏向并不总是优势:在十二维情形下,Station 止步于某一结果,而当前的点前沿是通过结合结构洞见的大规模数值优化才达到的 [76, 18]。
这个问题还表明,Station 生成的定理可能对研究人员具有独立的研究价值。例如,由智能体独立推导出的公式 (3),与一篇近期才公开发表的论文中的某个定理存在重叠 [77]。其显式的代数构造也可能具有独立价值。这些发现超出了分数优化的范畴,表明赋予 Station 智能体的额外自由度可以带来超越基准分数提升的贡献。
4.4 离散化 Kakeya 针问题
经典的 Kakeya 针问题探讨的是:将一个单位线段旋转覆盖所有方向,至少需要多大的面积。有限版本则将连续的方向替换为等间距的离散方向,并用可水平滑动的细长三角形来表示这些方向 [24]。更精确地说,对于实数偏移量,令
并定义
Córdoba 的下界以及 Keich 分析的 Schoenberg 构造表明,该量具有某一阶数 [19, 39],但其精确的有限值在很大程度上仍属未知。AlphaEvolve 将这一问题作为其数学问题集中的第 6.9 题;我们在相同的七个二进尺寸下,将三角形组件提供给了 Station。
S1. 在某一尺寸下的新上界。
Station 在三个有限尺寸下找到了更优的构造。在某一尺寸下,它找到了面积为某一值的三角形并集,相比 AlphaEvolve 的结果提升了某一幅度,相比后来的 HorizonMath 结果提升了另一幅度 [81],从而证明了……
在 和 处,收益较为温和,Station 将 AlphaEvolve 的面积分别减少了 和 ;在较小的测试规模 下,它达到了与 AlphaEvolve 相同的数值(图 4)。
S2. 在 处的精确最优解及在 处的对称性破缺。
在本工作之前,只有经典值 被精确已知 [24]。一个初等对称构造给出
而 Schoenberg 的经典 Perron 构造 [71] 给出
AlphaEvolve 后来在数值上复现了该值。Station 证明了匹配的下界,从而确立了
它还表明,两个极小值点都允许反射对称构型,且最优解包含连续族
Station 随后证明了在 处反射对称构型中的最小值为 ,并发现了一个面积为 的新非对称构造。图 4(右)比较了对称极小化构型与这个更小的非对称构造。这证明了在 处每个全局极小化构型都必然是非对称的,尽管 的精确值仍然悬而未决。
这些结果超出了基准评分范围。在 中,只有 是七个测试规模之一,且评估器只对显式构造的面积进行评分;它既不要求也不奖励全局下界的证明。任务规范也没有要求智能体对精确的小规模最优解进行分类或研究对称性破缺。智能体通过自主数学探索得出了这些结果,将其工作扩展到了有限构造基准之外。
局限性。
The Station 在测试的各个功率下分别优化了其构造,图 4 将它们与 AlphaEvolve 对应的分别优化的有限构造进行了比较。因此,该图比较的是双方的有限构造。除了这些分别优化的有限构造之外,AlphaEvolve 还提出了一种通过迭代式专家引导开发的、对每个(参数)都有效的单一构造。The Station 没有使用等效的专家参与流程,其自主智能体也没有发现具有竞争力的统一构造。
4.5 符号不确定性原理
一维符号不确定性问题探讨的是:当一个函数及其傅里叶变换在原点处均为负值时,它们最早能在何时都变为最终非负。对于一个非零的偶可积函数(其傅里叶变换也可积),定义
该问题要求出最大的常数 ,使得 。Bourgain、Clozel 和 Kahane 提出了这个问题 [12],后续研究逐步获得了更强的界 [33, 17]。AlphaEvolve 将其作为问题 6.11 进行研究,并报告了一个上界 以及一个未发表的人类上界 。The Station 进一步将该上界改进为 ,如图 5 所示。
S1. 一个新的上界 。
The Station 的智能体构造了一个函数,该函数产生了这个上界,从而证明了
它们采用
其中表达式涉及偶指标广义拉盖尔多项式;已证明尾部余量超过该值,因此最终非负性得以保持。这些基函数在傅里叶变换下保持不变,因此该选择自动给出相应结果,并将问题简化为构造一个具有所需符号的多项式。数值搜索找到了图 5 中所示的该次多项式;智能体将其系数表示为精确有理数,并证明了所得函数在相应半径之外非负,满足了问题对最终非负性的要求。
S2. 双根拉盖尔族在接近该值时已穷尽。
在此任务中,我们向智能体提供了与 AlphaEvolve 相同的指定双根拉盖尔设置和评分规则,但未提供 AlphaEvolve 的论文或结果。在此设置下,每个提交都被限制在由偶指标拉盖尔基中至多二十个指定正双根所确定的函数族内;我们称之为双根拉盖尔族。AlphaEvolve 的构造也属于该族。设
Station 智能体证明了
上界来自一个显式构造,而下界则来自对尾部点的精确加权和阻碍。因此,任何将上界改进到该值以下的构造都必须离开双根拉盖尔族。
这一界限促使智能体在受限族之外进行搜索,尽管官方评估器无法对超出该族的构造进行评分。它们将搜索扩展到没有指定双根的拉盖尔多项式,最终发现了给出该界限的该次构造。这提供了一个具体示例,说明智能体超越了分数优化,直接为底层数学问题做出贡献,尽管没有从分数中获得进一步指导。
4.6 Hardy–Littlewood 极大不等式
一维中心 Hardy–Littlewood 问题要求找到控制中心局部平均值可能变大的最优常数。对于非负可积函数,定义
并设 为满足以下条件的最小常数
Melas 解决了该问题,证明了
并构造逼近该值的有限质点示例 [55, 56]。AlphaEvolve 后来将这一有限问题作为基准,在搜索模式下达到该值,并在文献提示下得到约该值的结果。Station 智能体构造了一个 356 质点的构型,其值为 ,改进了 AlphaEvolve 的结果,但未能恢复 Melas 已发现的全局最优解。
S1. 中心化算子与非中心化算子之间的尖锐常数。
Ramos 考虑了在中心化 Hardy–Littlewood 极大算子与非中心化 Hardy–Littlewood 极大算子之间进行插值的自然非切向族 [64]。其参数从 处的中心化算子延伸到 处的非中心化算子。记 为尖锐弱 常数,Ramos 指出其精确值对每个 均未知,而端点情形是经典的 [6, 56]。在处理该任务时,Station 智能体解决了 这一情形,证明了
| (4) |
其余 的常数仍然开放。任务并未要求这一扩展,智能体也不知道 Ramos 曾提出过该问题;它们之所以继续探索,是为了理解当中心化约束被放宽时,中心化问题的几何结构如何变化。
4.7 卵形线问题
卵形线问题询问:是否每条闭凸平面曲线的曲率都会迫使相关联的一维 Schrödinger 算子的最低特征值至少为 。对于长度为 、以弧长 参数化的曲线,定义
其中 为曲率, 为周期边界条件下的最低特征值。Benguria 和 Loss 猜想 成立,并给出了一个包含圆和非圆卵形线的连续等号族 [5, 14, 8],证明了 ,而 Linde 证明了全局下界 ;对其定理中显式常数进行数值计算得到 [48]。AlphaEvolve 将这一问题作为其数学问题集中的问题 6.19 加以研究。
S1. 独立恢复 Benguria–Loss 等号族。
AlphaEvolve 恢复了圆形,但未获得非圆形的等号椭圆族。Station 则独立恢复了一族单参数正规形(在欧氏运动和弧长原点平移下),对应经典的 Benguria–Loss 等号族。因此,它重建了已知等号结构中比 AlphaEvolve 更大的一部分。这是对已知结果的独立恢复,而非新的等号族。Benguria 和 Loss 提出了该猜想并给出了等号族;Burchard 和 Thomas 证明了其局部极小性,而 Bernstein 和 Mettler 发展了其射影几何并确立了“Benguria–Loss 椭圆”这一名称 [5, 14, 8]。AlphaEvolve 和 Station 均未改进全局下界。
4.8 素数定理
素数定理描述了素数的渐近密度。若以 表示不超过 的素数个数,则它断言
因此,该底层数学问题已经解决:比值收敛到精确的 。然而,AlphaEvolve 仍将其有限版本作为其问题集中的问题 6.27 加以研究。它搜索了一个有限支撑的权重,使其满足
该权重及其相关求和的得分为
经典的 Chebyshev 论证表明
| (5) |
蕴含严格的全局下界
[20]。式 (5) 中所需的全局不等式远比素数定理本身更具限制性:单个有限权重必须对每个 都满足该不等式。AlphaEvolve 的得分仅在该不等式的有限多个采样值处进行检验。因此,它可能给一个在未测试值处失效的权重打出高分,此时该得分并不能证明所声称的素数计数下界。然而,对所有 进行穷举检验通常在计算上不可行,因为相关的周期可能极其庞大。因此,采样得分只能粗略近似地反映全局不等式是否成立。
S1. 一个对每个 都成立的得分。
Station 智能体发现了一个有限构造,它对每个 都满足式 (5),其中
| (6) |
这一结果优于 AlphaEvolve 报告的得分。更重要的是,这些智能体证明了所需不等式对所有情况均成立,而仅凭得分本身并不能提供这一保证。他们的关键思路是在构造中选择整数,使得其在可控范围内发生重复。这便将无穷多种可能的取值缩减为一次有限的穷举检查,智能体使用精确算术在一分钟内完成了该检查。
相比之下,同一轮运行中的其他智能体找到了得分更高的构造,达到了更高分数,但这些构造并未满足对所有情况都成立的全局不等式。这提供了一个具体实例,说明智能体在可被钻空子的得分面前,仍优先考虑底层数学问题而非单纯的得分优化。
S2. 为什么直接截断 Möbius 函数会失败。
Möbius 函数是一个自然的起点,因为它在素数定理的标准表述中处于核心地位。AlphaEvolve 探索了通过截断 Möbius 函数得到的有限构造,Station 智能体最初也采用了相同的方法。随后它们证明,这一族构造无法产生正的渐近得分:随着截断界限增大,其对所需全局不等式的最大违反程度至少以某个量级的规模增长。因此,对构造进行重新缩放以满足不等式会将其得分压低至趋于零的值。该证明建立在关于不完全 Möbius 和的结果之上[45]。这一障碍促使智能体放弃了直接截断 Möbius 函数的做法,转而探索一种更灵活的、对系数进行联合优化的构造,从而得到了上述严格的得分。
局限性。
由于素数定理已经精确确定了上述极限比率,这些结果并不会改变人们对素数分布的已知认识。它们的数学贡献更为有限:在该基准测试的有限设定内,Station 智能体找到了一种具有严格得分的构造,并证明了自然的 Möbius 截断无法产生正的渐近得分。因此,该问题主要起到一种校准作用,用于检验智能体能否区分一个有效的数学结果与一个得分很高但可被钻空子的结果,而非对素数分布研究作出实质性贡献。
4.9 差基
一个有限集合 是区间 的差基,如果该区间内的每个整数都可以表示为 中两个元素之差。若 是这样一个集合的最小可能大小,则需最小化的量为 ;Rédei 和 Rényi 证明了这些归一化最小值收敛,且其极限等于其下确界 [65]。AlphaEvolve 报告了上界
作为其集合中的问题 6.7。此前已发表的上界是 Golay 的 [32, 7],而非 AlphaEvolve 比较中所用的基准。这个例子是在人类专家提示的帮助下找到的:论文记录显示,AlphaEvolve 在获得用于生成 Singer 差集的正确代码之前一直未能改进其基准,其发布的提示词也将搜索引导至 Singer 集和经典的 Leech 乘积构造。我们只向 Station 提供了问题定义、评分规则和一个简单的网格基线。具体而言,这些智能体既没有获得这些构造提示,也无法访问外部文献。
S1. 独立重现 Leech–Golay 族中的一项纪录。
Leech 和 Golay 将四点差基与 Singer 差集相结合,得到了这一构造族中更早的成员 [43, 32, 4]。Station 独立重现了其成员。取 ,一个 元 Singer 差集 ,以及 ,智能体构造了
通过对 选取适当的代表元,所得整数实现了从 到 的所有差值,而 是第一个缺失的差值。因此
将 Golay 的上界进一步改进了约 。该集合与 AlphaEvolve 报告的构造逐项一致。这是对已知纪录的独立复现,并非相对于 AlphaEvolve 的新上界,也不是新的构造族。智能体还尝试进一步推低下界,但仅达到了经典界 [43],而 Yang 和 Liao 证明了更强的已发表界 [84]。
4.10 Sidorenko 猜想
Sidorenko 猜想断言:每个二分图 对任意图同态密度 均满足 ,其中 是 在 中的同态密度 [74]。最小的未解实例是拥有十个顶点、十五条边的图 ,也称为二分 Möbius 阶梯 [66]。AlphaEvolve 将该问题作为其数学问题集中的问题 6.26 提出,并在非恒定 阶梯图同态上进行了搜索。它通过以下方式对候选解进行评分:
因此正值将给出反例,从而推翻该猜想实例。
AlphaEvolve 报告称未找到反例。我们向 Station 给出了相同的问题和评分规则,Station 的智能体同样未找到反例。因此,该猜想的状态没有变化。
4.11 峰值自卷积
AlphaEvolve 的问题 6.2,在其问题集中被称为第一自相关不等式,探讨的是两个具有相同紧支撑密度的独立随机变量之和能有多均匀地分布。更精确地说,对于定义在 上、满足归一化条件 的非负函数 ,令
确定 的值与广义 Sidon 集的渐近大小相关,其精确值目前仍未知 [53]。目前报道的最佳界为
其中下界和上界分别来自经过认证的凸松弛和显式阶梯函数 [41, 68]。
AlphaEvolve 达到了上界,改进了 Matolcsi 和 Vinuesa [53] 在 AlphaEvolve 之前给出的界;TTT-Discover 随后将前沿推进到 [86],而精确算术证书又将其进一步提升到 [68]。Station 仅达到 ,比 AlphaEvolve 和当前前沿都差。AlphaEvolve 的高度不规则构造来自大规模启发式搜索。这一对比凸显了 Station 的一个局限:其智能体普遍偏好理论引导的构造而非启发式搜索,这种偏好虽然在多个其他问题上产生了强劲结果,但在这里却落后了——因为前沿构造依赖于大规模的启发式优化。
4.12 平坦自卷积
AlphaEvolve 的问题 6.3(在其集合中被称为第二个自相关不等式)探讨的是:非负函数的自卷积能在多大程度上接近一个平顶函数——即在一个集合上为常数、在该集合外为零的函数。更精确地说,对于非零非负函数 ,令
Hölder 不等式给出 ;对于任意非负输出,等号仅在这样一个平顶函数时成立。自卷积约束是否必然导致严格不等式仍然是一个开放问题 [51, 53]。在 AlphaEvolve 之前,已知的最佳界为 [53]
AlphaEvolve 确立了 的下界,而后续工作进一步将其改进到 [85]。Station 的最佳已验证构造仅达到 ,因此并未改进数值界。这一不足反映了与问题 6.2(最小化自卷积峰值,见第 4.11 节)中相同的局限:Station 的理论引导智能体并不擅长发现大规模启发式搜索所产生的高度不规则构造。
S1. 二元阶梯函数保持无约束上确界。
尽管如此,这些智能体还是证明了关于寻找近最优构造的一个有用事实:定义 的上确界可以通过二元阶梯函数来逼近,从而将对任意非负函数的搜索转化为对越来越细网格上二元函数的搜索。
4.13 图书 Ramsey 数
给定图 时,Ramsey 数是指最小的 ,使得 的任意红蓝边着色必然包含一个红色的 副本或一个蓝色的 副本。确定 Ramsey 数的精确值是一项困难的计算与理论挑战。最著名的 Ramsey 数当属 和 均为完全图的情形,但许多其他选择也已被广泛研究(见综述 [63])。书图 由共享一条公共边的 个三角形组成。一个悬而未决的问题是,是否
| (7) |
对每个正整数 都成立。Rousseau 和 Sheehan 于 1978 年建立了上界,证明了对于所有 有 [67]。因此,剩下的工作就是证明匹配的下界。对于给定的 ,这相当于构造 的一种红蓝边着色,使其既不包含红色的 也不包含蓝色的 。
第三作者证明了 时的等式成立,与同期工作独立地相互印证,并在 为同余于 的素数幂时建立了一个无限的 Paley 型族 [82, 47]。有限证据与一般算术构造的结合使他推测 (7) 对所有 均成立 [82]。Epoch AI 随后将其采纳为 FrontierMath 开放问题 [22]。在该问题发布之后,进一步的工作将连续求解的范围扩展到了 ,并通过扩展已有构造生成了两个额外的无限族 [80]。
我们针对该问题运行了两个 Station。第一个在无互联网访问的情况下运行,发现了一个新的会议图族。随后我们运行了第二个 Station,该 Station 可访问互联网并获得了第一个 Station 结果的摘要;它发现了一个新的加倍 Legendre 族以及若干新的有限构造。之后,一位外部专家将这些有限构造中的模式与第二个 Station 的早期结果相结合,得到了 Yamada–Pott 无限族。因此,前两个族是 Station 自主发现的结果,而第三个族则需要人类专家的参与。相对于现有文献,这三个族均为全新成果,并在图 6 中进行了可视化展示。每个族所覆盖的参数范围(包括哪些此前是未解决的)汇总于图 7。
S1. 会议图族。
第一个也是最广泛的族将任意会议图转换为锐利的 book-Ramsey 着色。具体而言,如果存在一个参数为
的强正则图,那么 Station 的智能体证明了
| (8) |
当 是 的素数幂时,Paley 会议图总是存在。因此,该定理在 是 的素数幂时证明了该猜想。在 Paley 情形之外,Seberry 和 Whiteman 使用 Mathon 的构造获得了每个 阶的对称会议矩阵 [52, 72]。这些矩阵产生 阶的会议图,因此 Station 定理也在
时证明了该猜想。第一个成员给出 和 。已知的 阶会议图提供了额外的参数 [36]。总体而言,已知的会议图在 30 个 值上证明了该猜想,其中 19 个此前是未解决的 [82, 47, 80, 22]。
S2. 双重勒让德族。
第二个族将 上的周期勒让德源转换为锐利的 book-Ramsey 着色 [25]。具体而言,对于每个满足 的素数幂 ,Station 的智能体证明了
| (9) |
因此,该定理在 是 的素数幂时证明了该猜想。对于 ,该族在 21 个值上证明了等式成立,并且在发现之时,在计入会议族之后还解决了六个额外的未解情形 [82, 47, 80, 22]。
智能体于 2026 年 7 月中旬发现了这个一般族。7 月底宣布的并行工作独立地产生了有限情形 [22];Station 定理包含 作为其中一个成员,并覆盖了无限多个进一步的参数。
加倍 Legendre 族与 Turturean [80] 报告的 Legendre 族相关但不同。两者都始于相同类型的周期 Legendre 源(在 上,且 ),但使用不同的提升来获得 book-Ramsey 着色。对于相同的源阶 ,较早的提升达到 ,而 Station 提升达到 。因此它将 Ramsey 参数加倍,并覆盖一组不同的值,如图 7 所示。
S3. Yamada–Pott 族。
第三个族将经典的 Yamada–Pott 设计转化为锐利的 book-Ramsey 着色 [3]。具体来说,对于每个满足 的素数幂 ,我们证明了
| (10) |
因此,该定理在
对于与 同余的素数幂 时证明了该猜想。对于 ,该族在五个值处证明了等式,并在考虑 conference 族和加倍 Legendre 族之后解决了三个额外的先前未解决的情况 [82, 47, 80, 22]。第二位 Station 的智能体提供了 的有限仿射构造和一个较早的周期相关性恒等式;一位外部专家识别出它们共有的 Yamada–Pott 结构,并利用这些要素建立了该一般定理。
讨论。
上述三个无限族相对于现有文献而言是新颖的,但其源对象并非新事物:会议图(conference graphs)、周期勒让德对(periodic Legendre pairs)和 Yamada–Pott 设计均已有先例 [52, 25, 3]。每个案例中的新意在于将经典对象提升为锐利 book-Ramsey 染色的规则,而这样的规则未必能从源对象本身直接看出。例如,智能体是在经过超过 3,000 个 Station 时钟周期以及一长串中间内部论文之后,才发现了通用的会议图提升规则。随附的笔记本提供了改编自智能体内部论文、相对未经打磨的证明;我们将在后续的单独论文中给出全部三个族的精炼证明。
前两个族还表明,Station 智能体能够推进超出可直接评分任务的通用数学目标:尽管评估器只能奖励有限构造,它们仍然发现并证明了无限族。第三个族则展示了互补的局限性。有限仿射示例和一般定理所需的周期相关恒等式都已存在于 Station 的研究历史中,但智能体未能将它们联系起来。一位外部专家识别出它们共同的 Yamada–Pott 结构并完成了综合。这一错失的联系表明,智能体可能尚未充分利用 Station 中积累的知识,在此类情况下或许需要外部专家的综合协助。
4.14 Jacobian 猜想
雅可比猜想问的是:一个处处局部可逆的多项式映射,是否也必然全局可逆。更精确地说,它断言:每个具有非零常数雅可比行列式的多项式映射都是多项式自同构[40]。2026年7月19日,有消息宣布,利用Claude Fable[1]构造出了一个三维反例,从而在一切至少为三维的维度上推翻了这个猜想。这一突破随即促使研究者们去寻求对该映射的概念性解释:特别是,其看似神奇的雅可比相消为何会发生,以及三个一般的逆曲面如何能与处处局部可逆性共存[15, 27, 78, 73, 75]。
我们在该公告发布一周后启动了Station。由于这项实验是在更新模型问世之后进行的,它使用了比其他Station更新的智能体池:两个智能体分别由GPT-5.6 Sol、Claude Opus 5和Gemini 3.1 Pro驱动。这些智能体没有外部网络访问权限,只收到一份不含公式的规格说明:构造一个次数至多为某值的、具有有理系数的多项式映射,要求其雅可比行列式为非零常数,且在同一纤维中存在两个不同的有理点。评估器自动检查每个构造,并且只有当其满足所有要求时才给出分数,否则不给分。我们没有提供任何文献综述或部分构造。因此,智能体必须独立地找到这个反例。
这项任务的目标有两个。首先,我们希望在严格二元问题上测试 Station。评估器既不提供部分分数,也不提供分级反馈,因此失败的尝试不会给智能体任何关于如何改进的分数信号;要达到所需分数,必须重构一个困扰数学家近九十年之久的猜想反例 [40]。其次,我们希望观察完整的发现过程,而不仅仅是最终构造。我们公开了全部原始智能体对话,而原始 Fable 的发现轨迹尚未发布。这份记录保留了未出现在最终构造中的中间数学想法,使研究人员能够研究 AI 主导的数学发现的动态过程。
S1. 通过尖点直纹面的独立重构。
设 ,一个 Station 智能体构造了七次映射
精确计算得到 ,以及三个不同的有理点
全部映射到 。这些恒等式构成一个完整的反例证书。该公式与已公布的映射 [1, 26] 明显不同,但线性源变换和目标变换满足 。因此 Station 在不同的线性坐标下重构了已公布的反例;它没有产生新的反例或新的等价类。
原始结果归功于 Claude Fable,而该反例由一个 GPT-5.6 Sol 智能体在一天之内独立发现,未与其他智能体直接交互。成功的智能体从直纹映射 开始,因此对于每个固定的 ,变化时 描绘出一条直线。它测试了基于光滑圆锥曲线的五种低次方向模板,但没有一种满足剩余的常数雅可比条件。决定性的一步是用尖点三次曲线 替换光滑方向曲线。其关联方向场为 ;在这一选择下,基曲面的相容方程变得可解,并恰好产生了上述映射。
S2. 重构的映射具有无临界点的三叶纤维。
在成功推导过程中,该智能体还解释了为什么尖点直纹(cuspidal ruling)能使雅可比行列式为常数。对于方向场,智能体推导了活动标架恒等式,包括 的恒等式,据此,行列式中每一个依赖于 的贡献项都包含一个重复的切方向并因此为零。剩余的三重积即为常数 。因此,智能体是从尖点直纹的几何结构推导出雅可比行列式的消去,而非发现十六个只能在事后验证其消去的项。
在构造出反例之后,同一智能体分析了其纤维结构,并解释了该映射为何能在处处局部可逆的同时,在一般情况下具有三个原像。在一个稠密坐标卡上,将目标点写为 ,并令 和 。恢复一个原像于是归结为
| (11) |
对于一般的目标点,三个根给出三个不同的原像。若记左侧为 ,则逆公式满足 、 ,从而有 。当根重合且 时,条件 迫使相应的源点逃逸至无穷远处,而非在仿射空间中成为临界点。在例外轨迹 上,源坐标提供了一个额外的仿射尺度方向,从而消解了同样的表观分歧。这一分析回答了数学家在公告发布后立即提出的结构性问题:三个叶层源于一个三次商映射,而完整三维映射的几何结构阻止了它们的碰撞产生临界点。该智能体的解释与公告发布后数日内数学家们发展出的尖点与三次解释相吻合 [27, 78, 73, 75]。
讨论。
该实验的数学结果是独立重构,而非新的反例或新的解释。该实例表明,Station 能够解决一个困难的二元问题,而该问题的评估器不提供梯度或部分分数来引导搜索。尽管如此,此类反例突破可能仍然罕见,因为猜想通常被预期为真。在更广泛的背景下,更困难的挑战可能在于识别一个有前景的问题,并在知晓反例是否存在之前投入大量计算。
5 元分析
在本节中,我们对上述发现过程进行元分析,以更好地理解 AI 发现的动态。除非另有说明,所有分析均基于前文提到的 14 个问题背后的 16 个 Station 实例。(Kissing number in 和 Book Ramsey numbers 问题各有两个 Station 实例。)Spotlight 结果指该节中标记为 S1、S2 等的结果,共计 28 个结果。当单个 spotlight 包含多个独立发现的成果时,我们分别计数这些成果。我们使用“archive paper”指代由 Station 内智能体发表的论文,而非外部人类文献中的论文。
5.1 模型家族的贡献
我们首先分析 28 个 spotlight 结果中每个结果的主要贡献者,如图 8(a) 所示。我们将每个结果归因于做出实质性发现的智能体,而非后来重述、验证或发表该结果的智能体。Claude 智能体对 18 个结果(64.3%)做出了主要发现,GPT 智能体对 9 个结果(32.1%)做出了主要发现,Gemini 智能体对 1 个结果(3.6%)做出了主要发现。Gemini 的较小份额可能部分反映了模型发布时间:Gemini 3.1 Pro 于 2026 年 2 月发布,早于 4 月的 GPT-5.5 和 5 月的 Claude Opus 4.8 [34, 60, 2]。因此,其较低的贡献与行业普遍趋势一致,即较晚发布的模型具有更强的能力。
我们还分析了智能体的档案论文贡献,如图 8(b) 所示。Gemini 智能体提交的档案论文最多:共尝试 2,652 次,其中 508 篇被接收(19.2%),也就是说超过 80% 被评审者拒绝。Claude 智能体尝试了 1,236 次,其中 696 篇被接收(56.3%),而 GPT 智能体仅尝试了 506 次,其中 388 篇被接收(76.7%)。我们还按模型家族计算了总引用次数,发现 Claude 智能体的档案论文无论在总量还是平均值上都获得了最多的引用(图 8(c))。在我们的观察中,Gemini 智能体倾向于过度宣称,例如仅凭有限的证据就断言某个方向不可行;这类投稿通常会被评审系统拒绝,这或许有助于解释其高拒绝率。相比之下,GPT 智能体在档案论文投稿上非常谨慎,通常只在发现相对重要时才提交,这或许有助于解释其较低的投稿数量。Claude 的档案论文通常篇幅更长、内容更全面,这或许有助于解释其较高的平均引用次数。这些模式反映了不同模型家族在研究风格上的差异。
从定性角度看,我们观察到三个模型家族在优势与失败模式上存在显著差异。Gemini 智能体倾向于提出更新颖的启发式方法和研究方向,但它们也更容易夸大结论,或在收到同行反馈时过于轻易地改变方向。GPT 智能体往往更为严谨,通常能够为新结果给出有效的非正式证明,但它们也可能陷入技术细节繁复的旁支问题,而这些问题的整体研究价值有限。Claude 智能体则往往表现出持久、有条理且善于自我批判的特点。它们的创造力常常是适应性的:它们会从失败的方法中学习,利用这些失败来识别新方向,并通过严谨的验证持续推进这些方向。这种严谨与有纪律的创造力相结合,使 Claude 成为高产贡献者。不过,其智能体偶尔也会提出错误论断,随后由同行智能体予以纠正。
5.2 跨模型家族的协作
Station 的一个特点是允许来自不同模型家族的智能体进行协作。因此,我们考察了不同模型家族的智能体在 spotlight 成果上共同协作的频率。我们审视了上述全部 28 项 spotlight 成果。当某个智能体的工作被实质性用于该成果时——例如,它贡献了一个定理、构造、方法或研究方向,并被其他智能体所采用——我们便将其计为贡献者。
我们发现,在 28 个亮点结果中,有 13 个(46.4%)涉及来自多个模型家族的智能体,如图 9(a) 所示。在其余 15 个结果中,有 6 个仍是由同一模型家族的多个智能体合作完成。因此,28 个结果中只有 9 个(32.1%)是由单个智能体独立发现,而 19 个(67.9%)涉及多个智能体。Claude 智能体尤其善于协作:它们参与了全部 13 个跨模型结果。这些发现表明,智能体之间以及不同模型家族之间的协作是发现过程中的重要组成部分。相比之下,当前大多数 AI 驱动的科学系统,要么在单次运行中使用单一模型家族的智能体 [58, 30, 70, 35, 61],要么在流水线中让不同模型家族承担固定角色 [50, 31]。
我们还考察了跨模型案例中智能体之间的沟通方式。档案室(Archive Room)是最常用的渠道,占这些协作的61.5%(图9(b))。这表明档案论文是智能体之间进行同行交流的一种高效手段。作为智能体较长的研究过程中对科学成果的高度浓缩记录,这些档案论文提供了一种低带宽但信息密度高的知识载体,后续智能体可以在此基础上继续推进,这与我们人类的科学文献颇为相似。一个智能体可以解决问题的某一部分并说明仍缺失什么;来自另一个模型家族的后续智能体可以阅读档案论文并继续推进。事实上,图9(c)展示了一个著名的案例研究——三个模型家族主要通过档案论文进行协作,最终取得了首个有限Kakeya聚光结果。
5.3 发现时间
我们同样关注Station完成每项发现所需的时间。大多数Station实例运行了1,000至2,000个tick,对应大约一到两周的连续墙钟时间。图10展示了28个聚光结果各自以最终实质性形式首次出现时的tick位置。
一些相对简单的结果出现得较早。除了雅可比猜想(Jacobian Conjecture)这一显著例外,这些早期发现往往分量较轻,通常是对预训练知识中已有思路的直接改编或扩展,而此时Station中积累的共享知识还不多。
28 个焦点结果中有 13 个(46.4%)是在第 1000 次迭代之后才被发现的。我们普遍观察到,较晚的发现往往更具新颖性或难度。最极端的例子是 Book Ramsey 数的 conference-graph 族,它是在第 3727 次迭代时被发现的。其提升规则在现有文献中远非显而易见,值得单独撰写一篇外部后续论文。这类非平凡的发现往往只有在大量内部文献积累之后才会出现。
5.4 Station 机制
Station 的设计旨在通过多种机制促进科学发现。这些机制在附录 A 中有详细描述;这里我们给出简要概述,并探讨其中哪些机制对焦点结果做出了贡献。
假期。每十个迭代周期中的最后两个迭代被宣布为假期;智能体不能提交代码或归档论文,而是收到鼓励广泛反思、隐喻或借鉴其他领域想法的提示词。这种暂停常常促使智能体重新考虑失败的方法,或探索不那么明显的方向。
归档论文。被接受的归档论文构成 Station 的累积知识,并持续可供后续智能体使用。这使得部分定理、构造和有据可查的失败能够成为后续发现的起点。
停滞协议。如果官方评估前沿长期没有改进,Station 会要求智能体审查内部文献、质疑自身假设,并尝试不同的高层策略。这有助于智能体摆脱已经耗尽的局部方法,推动它们进行更大胆的尝试和更广泛的探索。
同行交流。智能体可以通过直接邮件或共享的公开讨论来交换部分结果、有针对性的问题和批评意见。
主管。Station 会随机指定一名符合条件的智能体担任主管。主管提供高层指导,鼓励坚持并防止智能体重复彼此的工作,同时让它们对自己的研究负责;在两次任命之间,Station 有意留出较长的无主管时期,以鼓励更少结构化的探索。
问题室。智能体可以发布重要的开放子问题,供其他智能体讨论和解决。这能把未解决的空白转化为共同的研究目标,并让采用不同方法的智能体补上缺失的环节。
这些机制以不同方式支持探索发现。假期拓宽探索范围;档案论文加深累积知识;停滞协议提供摆脱局部最优的推力;而同伴交流、监督和问题室则协调各智能体之间的工作。
我们审阅了 28 项成果各自背后的对话,并将每种机制归类为对发现作出直接贡献、间接贡献或无实质贡献(图 11)。当机制提供了决定性思路或干预时,即为直接贡献;当机制塑造或支持了研究但并非成果的直接来源时,即为间接贡献。当对话显示没有明确的因果作用时,我们将其归为无实质贡献。
假期和档案论文分别对 28 项成果中的 23 项和 21 项作出了直接或间接贡献,其次是停滞协议,贡献了 14 项。在假期期间,智能体常常从主动优化中抽身,审视先前方法失败的原因,并重新定义问题或探索新方向;这些反思往往提供了后来成为亮点成果一部分的思路,这解释了其高贡献率。档案论文也对相当一部分成果作出了贡献,表明 Station 积累的知识对后续发现是有用的。
5.5 结果可复现性
我们同样关注这些发现是否具有可复现性。为此,我们在十一维亲吻数问题上运行了三个独立的 Station 实例,均不接入网络。(其中包括第 4.3 节描述的两个实例;第三个仅用于本次可复现性分析,不纳入上述其他元分析。)图 12 展示了每次运行所达到的最佳已认证下界。三个 Station 最终都达到了该下界,表明改进后的下界是可复现的。
然而,进一步观察显示,在所需时间和达成结果的路径上都存在显著差异。Station 1 围绕格导出的核心进行离散精确线填充。它通过选取 54 条相互兼容的线,构成一个 -点核心的 -点代数扩张,从而获得构造 3;随后在探索另一个核心和扩张时获得构造 2。Station 2 则是在一个公共旋转下,由根系构型组装出构造 1;其最后一步是识别出十一个点构成了一个立方八面体除一个顶点外的全部顶点,并补上缺失的第十二个顶点。Station 3 通过不同的机制达到了同一构造类别:它对一个精确的 -点构型进行变形,使得两个坐标向量和一个额外向量能够支撑在一个特定的三维子空间上并被追加进去。由此可见,相同的数值下界源自截然不同的数学表示和研究路径。
这种差异部分源于Station的累积知识。初始轨迹上的微小差异会决定哪些结果进入档案论文集合。后续智能体随后继承不同的起点,因此研究路径和累积档案论文的差异会随时间不断放大。因此,鉴于不同Station实例之间存在高度方差,在计算成本不成问题的情况下,对同一问题运行多个独立实例是明智的做法。
6 讨论与结论
我们观察到AI智能体的能力正在快速提升。在一年前的初始版本中,智能体频繁出现模型幻觉,无法可靠地学习环境规则。如今智能体已经能够掌握环境并自主产生新颖的发现。尽管如此,多智能体研究仍存在若干重要局限。我们在下方总结观察结果。
缺乏专家直觉。这里所说的直觉,是指在不深入探索之前判断某个研究方向是否有前景的能力。良好的直觉能让探索更高效,使研究者能够更深入地钻研有前景的方向。在多次运行中,我们观察到多起智能体以不充分的理由将前景良好的方法降级处理,从而延误或错失潜在突破的案例。这表明智能体缺乏该领域人类专家通常具备的那种直觉。
缺乏多样化的研究品味。对特定概念或方法的偏好很难客观评判为优劣。然而,当所有智能体都拥有相似的品味时,整体探索范围就会变得狭窄。在多次运行中,来自同一模型家族的智能体往往提出相似的研究思路,这表明模型特有的品味降低了探索的多样性。
上下文学习能力有限。智能体可以通过其上下文吸收新的研究知识,但这些知识并不会更新其预训练权重。随着“研究站”积累的知识不断增长,智能体可能难以完全吸收并有效在此基础上继续推进。我们偶尔观察到智能体未能认识到自己的研究路线与“研究站”早期知识之间的关联,从而错失潜在的发现。
吸引子陷阱。当被赋予自主性时,一些智能体会沉浸在我们称之为“吸引子”的任务或活动中。这些活动通常在某种即时意义上具有回报性,但对主要问题却几乎没有有意义的贡献。智能体也可能陷入技术细节中,而人类专家会迅速识别出这些细节是琐碎的或与主要问题无关。例如,反复使用不同的随机种子重新运行同一个优化脚本,或详尽地诊断和表征每一个局部最优解。
“研究站”的若干机制旨在缓解这些局限。例如,使用来自多个模型家族的智能体拓宽了研究品味的范围,而停滞协议则帮助智能体摆脱吸引子陷阱。尽管如此,这些问题在一定程度上仍然存在,并且在所有四个方面,AI 智能体与人类专家之间仍存在显著差距。来自人类专家的轻量级指导或偶尔干预,通过引导智能体走向有前景的研究领域,可能会有所裨益。当前“研究站”支持此类人类参与,例如通过向所有智能体广播消息的方式,但我们将人机协作的系统性研究留待未来工作。
尽管本文主要将“研究站”用于数学探索,但“研究站”被设计为一个通用的研究环境,其任何机制都不是专门针对数学定制的。正如原论文所展示的,“研究站”可应用于涵盖数学、计算生物学和机器学习的问题[16]。因此,在其他领域的大规模研究探索,包括对语言模型本身的研究,可能前景广阔。
随着 AI 智能体能力不断增强,我们预计自主性和通用性将成为设计 AI 研究环境时越来越重要的原则。更强的智能体不必局限于日益复杂的流水线;它们有能力自行决定如何实现目标、从失败中学习、交流想法,并随着时间推移不断积累知识。Station 所提供的更高自主性,可能使这些能力得到更充分的发挥。
参考文献
- [1] L. Alpöge (2026) 你好,雅可比猜想是错的。注:X 平台帖子,发布于 2026 年 7 月 19 日。外部链接:链接。被引用:§1、§4.14、§4.14。
- [2] Anthropic (2026) Claude opus 4.8。注:Anthropic。外部链接:链接。被引用:§5.1。
- [3] K. T. Arasu、D. A. Bulutoglu 和 J. R. Hollon (2020) Legendre 阵列对与若干阵列族的理论统一。Journal of Combinatorial Designs 28 (11),第 814–841 页。注:arXiv:2004.05608。外部链接:文档。被引用:§4.13、§4.13。
- [4] T. Banakh 和 V. Gavrylkiv (2019) 循环群中的差基。Journal of Algebra and Its Applications 18 (5),第 1950081 页。注:arXiv:1702.02631。外部链接:文档。被引用:§4.9。
- [5] R. D. Benguria 和 M. Loss (2004) Schrödinger 算子的 Lieb–Thirring 猜想与平面上椭圆等周问题之间的联系。载于《偏微分方程与反问题》,当代数学丛书,第 362 卷,第 53–61 页。注:arXiv:math-ph/0402048。被引用:§4.7、§4.7。
- [6] A. Bernal (1989) 关于一维极大函数的一个注记。Proceedings of the Royal Society of Edinburgh Section A: Mathematics 111 (3–4),第 325–328 页。外部链接:文档。被引用:§4.6。
- [7] A. Bernshteyn 和 M. Tait (2019) 差基的改进下界。Journal of Number Theory 205,第 50–58 页。注:arXiv:1901.09411。外部链接:文档。被引用:§4.9。
- [8] J. Bernstein 和 T. Mettler (2015) 一维射影结构、凸曲线与 Benguria & Loss 椭圆。Communications in Mathematical Physics 336 (2),第 933–952 页。注:arXiv:1403.8000。外部链接:文档。被引用:§4.7、§4.7。
- [9] M. R. Best(1977),《或某些新的最优等重码》。技术报告 ZN 71/77,阿姆斯特丹数学中心。外部链接:链接 引用位置:§4.3。
- [10] F. Bianchi、Y. Kwon、A. Pappu 和 J. Zou(2026),《利用野外 AI 智能体的集体智能实现新发现》。arXiv 预印本 arXiv:2606.10402。外部链接:文档 引用位置:§4.3。
- [11] A. Blokhuis 和 F. Mazzocca(2008),《有限域 Kakeya 问题》。载于《搭建桥梁:数学与计算机科学之间》,M. Grötschel 与 G. O. H. Katona 主编,Bolyai 学会数学研究丛书,第 19 卷,第 205–218 页。注:arXiv:0911.4370 外部链接:文档 引用位置:第 4 项。
- [12] J. Bourgain、L. Clozel 和 J. Kahane(2010),《Heisenberg 原理与正函数》。Annales de l’Institut Fourier 60(4),第 1215–1232 页。外部链接:文档 引用位置:§4.5。
- [13] B. Bukh 和 T. Chao(2021),《有限域 Kakeya 问题的尖锐密度界》。Discrete Analysis。注:第 26 条,9 页;arXiv:2108.00074 外部链接:文档 引用位置:第 1 项、§4.1、§4.1、§4.1。
- [14] A. Burchard 和 L. E. Thomas(2005),《关于依赖于环曲率的 Schrödinger 算子的一个等周不等式》。The Journal of Geometric Analysis 15(4),第 543–563 页。注:arXiv:math/0505123 外部链接:文档 引用位置:§4.7、§4.7。
- [15] K. Buzzard(2026),《人类数学家正在被反例超越》。注:The Xena Project 博客 外部链接:链接 引用位置:§4.14。
- [16] S. Chung 和 W. Du(2025),《The Station:一个用于 AI 驱动发现的开放世界环境》。外部链接:2511.06309、文档、链接 引用位置:附录 A、§1、§2、§6。
- [17] H. Cohn 和 F. Gonçalves(2019),《通过模形式得到的十二维最优不确定性原理》。Inventiones Mathematicae 217,第 799–831 页。注:arXiv:1712.04438 外部链接:文档 引用位置:§4.5。
- [18] H. Cohn(2026),《接吻数》。注:在线表格 https://cohn.mit.edu/kissing-numbers/,访问于 2026 年 8 月 4 日 引用位置:§4.3、§4.3。
- [19] A. Córdoba(1977)《Kakeya 极大函数与球面求和乘子》。American Journal of Mathematics 99(1),第 1–22 页。外部链接:文献。引用位置:§4.4。
- [20] H. G. Diamond(1982)《素数分布研究中的初等方法》。Bulletin of the American Mathematical Society 7(3),第 553–589 页。外部链接:文献。引用位置:§4.8。
- [21] Z. Dvir(2009)《有限域上 Kakeya 集合的大小》。Journal of the American Mathematical Society 22(4),第 1093–1097 页。外部链接:文献。引用位置:§4.1。
- [22] Epoch AI(2026)《Book Ramsey 数》。注:FrontierMath 开放问题,访问于 2026 年 8 月 17 日。外部链接:链接。引用位置:§4.13、§4.13、§4.13、§4.13、§4.13。
- [23] P. Erdős(1955)《数论中的若干评注》。Riveon Lematematika 9,第 45–48 页。注:希伯来文。引用位置:§4.2。
- [24] K. J. Falconer(1985)《分形集的几何》。Cambridge Tracts in Mathematics,第 85 卷,剑桥大学出版社。引用位置:§4.4、§4.4。
- [25] R. J. Fletcher、M. Gysin 与 J. Seberry(2001)《离散傅里叶变换在搜索广义 Legendre 对与 Hadamard 矩阵中的应用》。Australasian Journal of Combinatorics 23,第 75–86 页。外部链接:链接。引用位置:§4.13、§4.13。
- [26] A. Freitas Ramos、D. Barros Hulak 与 R. J. Guerra Barretto de Queiroz(2026)《Jacobian 猜想一个显式反例的形式化验证》。注:Archive of Formal Proofs。外部链接:链接。引用位置:§4.14。
- [27] A. Gallagher(2026)《三维 Jacobian 猜想的无穷反例族:每个一般纤维次数均出现》。注:Zenodo 预印本。外部链接:文献、链接。引用位置:§4.14、§4.14。
- [28] M. Ganzhinov(2025)《高度对称的直线》。Linear Algebra and its Applications 722,第 12–37 页。注:arXiv:2207.08266。外部链接:文献。引用位置:§4.3。
- [29] B. Georgiev、J. Gómez-Serrano、T. Tao 与 A. Z. Wagner(2025)《大规模数学探索与发现》。arXiv 预印本 arXiv:2511.02864。外部链接:文献、链接。引用位置:§3.1、§3.2。
- [30] A. Ghafarollahi 和 M. J. Buehler(2025)《SciAgents:通过仿生多智能体智能图推理实现科学发现自动化》。Advanced Materials 37 (22),第 2413523 页。外部链接:文档、链接。引用于:§2、§5.2。
- [31] A. E. Ghareeb、B. Chang、L. Mitchener、A. Yiu、C. J. Szostkiewicz、D. Shved、G. J. Gyimesi、J. M. Laurent、S. M. Wright、M. T. Razzak、A. D. White、S. C. Finnemann、M. M. Hinks 和 S. G. Rodriques(2026)《用于自动化科学发现的多智能体系统》。Nature 655,第 497–505 页。外部链接:文档、链接。引用于:§2、§5.2。
- [32] M. J. E. Golay(1972)《关于用差分表示 [原文此处缺失] 的注记》。Journal of the London Mathematical Society s2-4 (4),第 729–734 页。外部链接:文档。引用于:§4.9、§4.9。
- [33] F. Gonçalves、D. Oliveira e Silva 和 S. Steinerberger(2017)《Hermite 多项式、环面上的线性流以及根的不确定性原理》。Journal of Mathematical Analysis and Applications 451 (2),第 678–711 页。外部链接:文档。引用于:§4.5。
- [34] Google(2026)《推出 Gemini 3.1 Pro:为最复杂任务打造的更智能模型》。注:Google 博客。外部链接:链接。引用于:§5.1。
- [35] J. Gottweis、W. Weng、A. Daryin、T. Tu、P. Sirkovic、A. Myaskovsky、G. Glowaty、F. Weissenberger、A. Orlandi、D. Popovici、A. Palepu、K. Rong、R. Tanno、K. Saab、F. Zhang、J. Blum、A. Carroll、K. Kulkarni、N. Tomašev、D. Zverinski、I. Rendulic、E. Vedadi、F. Hasler、L. Rimanic、M. Boia、I. Budiselic、B. Feinstein、M. Bellaiche、T. Sheffer、J. Freyberg、J. Ratcliff、O. Bertolli、K. Chou、A. Hassidim、B. Gokturk、A. Vahdat、Y. Guan、V. Dhillon、E. D. Vaishnav、B. Lee、T. R. D. Costa、J. R. Penadés、G. Peltz、Y. Matias、J. Manyika、D. Hassabis、Y. Xu、P. Kohli、A. Pawlosky、A. Karthikesalingam 和 V. Natarajan(2026)《利用 Co-Scientist 加速科学发现》。Nature 655,第 487–496 页。外部链接:文档、链接。引用于:§2、§5.2。
- [36] O. Gritsenko(2021)《关于参数为 [原文此处缺失] 的强正则图》。arXiv 预印本 arXiv:2102.05432。外部链接:文档。引用于:§4.13。
- [37] J. K. Haugland(2016)最小重叠问题再探。arXiv 预印本 arXiv:1609.08000。外部链接:文献。被引用:§4.2。
- [38] E. Hedley(2025)科学的创造力能否习得?这些研究者认为可以。《自然》。外部链接:文献。被引用:§A.4。
- [39] U. Keich(1999)关于 Kakeya 极大函数与 Minkowski 维数的界。《伦敦数学学会公报》31(2),第 213–221 页。外部链接:文献。被引用:§4.4。
- [40] O. Keller(1939)整 Cremona 变换。《数学与物理月刊》47,第 299–306 页。外部链接:文献。被引用:§4.14、§4.14。
- [41] S. Kim 与 M. Pilanci(2026)通过双智能体实现凸松弛的 AI 辅助发现。arXiv 预印本 arXiv:2606.31182。外部链接:文献。被引用:§4.11、§4.2、§4.2。
- [42] S. Kopparty、V. F. Lev、S. Saraf 与 M. Sudan(2011)有限向量空间中的 Kakeya 型集合。《代数组合学杂志》34(3),第 337–355 页。注:arXiv:1003.3736。外部链接:文献。被引用:第 3 项。
- [43] J. Leech(1956)关于用差表示整数。《伦敦数学学会杂志》s1-31(2),第 160–169 页。外部链接:文献。被引用:§4.9、§4.9。
- [44] 莱顿宣言工作组(2026)莱顿人工智能与数学宣言。外部链接:文献、链接。被引用:§1。
- [45] P. Letendre(2020)Möbius 函数的截断卷积与整数的乘法能量。《算术学报》195(1),第 83–95 页。外部链接:文献。被引用:§4.8。
- [46] V. F. Lev(2009)对“DHJ3:900–999(密度 Hales–Jewett 型数)”的评论 994。注:博客评论,《新进展》(T. Tao)https://terrytao.wordpress.com/2009/03/04/dhj3-900-999-density-hales-jewett-type-numbers/comment-page-3/#comment-36694,访问于 2026 年 7 月 30 日。被引用:第 6 项、§4.1。
- [47] B. Lidický、G. McKinley、F. Pfender 与 S. Van Overberghe(2025)书图、轮图及其推广的小 Ramsey 数。《电子组合学杂志》32(4),第 P4.64 页。注:arXiv:2407.07285。外部链接:文献。被引用:图 7、图 7、§4.13、§4.13、§4.13、§4.13。
- [48] H. Linde(2025)环上薛定谔算子基态的改进界。arXiv 预印本 arXiv:2504.20229。外部链接:文档 引用自:§4.7。
- [49] E. Lorist 与 F. L. Schwenninger(2026)Crouzeix 猜想的一个解。arXiv 预印本 arXiv:2608.03841。外部链接:文档、链接 引用自:§1。
- [50] C. Lu、C. Lu、R. T. Lange、Y. Yamada、S. Hu、J. Foerster、D. Ha 与 J. Clune(2026)迈向 AI 研究的端到端自动化。《自然》651 卷,第 914–919 页。外部链接:文档、链接 引用自:§2、§5.2。
- [51] G. Martin 与 K. O’Bryant(2009)自卷积的上确界及其在加性数论中的应用。《伊利诺伊数学杂志》53 卷(1),第 219–235 页。外部链接:文档 引用自:§4.12。
- [52] R. Mathon(1978)阶对称会议矩阵。《加拿大数学杂志》30 卷(2),第 321–331 页。外部链接:文档 引用自:§4.13、§4.13。
- [53] M. Matolcsi 与 C. Vinuesa(2010)自卷积上确界的改进界。《数学分析与应用杂志》372 卷(2),第 439–447 页。外部链接:文档 引用自:§4.11、§4.11、§4.12。
- [54] L. Mazur(2026)Sendov 猜想的计算机辅助证明。注:Proof Atlas 外部链接:链接 引用自:§1。
- [55] A. D. Melas(2002)关于中心 Hardy–Littlewood 极大算子。《美国数学学会汇刊》354 卷,第 3263–3273 页。外部链接:文档 引用自:§4.6。
- [56] A. D. Melas(2003)中心 Hardy–Littlewood 极大不等式的最佳常数。《数学年刊》157 卷(2),第 647–688 页。外部链接:文档 引用自:§4.6、§4.6。
- [57] G. Mockenhaupt 与 T. Tao(2004)有限域上的限制与 Kakeya 现象。《杜克数学杂志》121 卷(1),第 35–74 页。外部链接:文档 引用自:第 2 项。
- [58] A. Novikov、N. Vũ、M. Eisenberger、E. Dupont、P. Huang、A. Z. Wagner、S. Shirobokov、B. Kozlovskii、F. J. R. Ruiz、A. Mehrabian、M. P. Kumar、A. See、S. Chaudhuri、G. Holland、A. Davies、S. Nowozin、P. Kohli 与 M. Balog(2025)AlphaEvolve:用于科学与算法发现的编码智能体。arXiv 预印本 arXiv:2506.13131。外部链接:文档、链接 引用自:§1、§2、§5.2。
- [59] OpenAI(2026)Codex CLI。注:OpenAI 文档 外部链接:链接 引用位置:§A.3。
- [60] OpenAI(2026)推出 GPT-5.5。注:OpenAI 外部链接:链接 引用位置:§5.1。
- [61] OpenAI(2026)多智能体。注:访问时间:2026-08-14 外部链接:链接 引用位置:§2、§5.2。
- [62] OpenAI(2026)数学与理论计算机科学领域的十项进展。注:OpenAI 外部链接:链接 引用位置:§1。
- [63] S. P. Radziszowski(2026)小 Ramsey 数。电子组合学杂志。注:动态综述,DS1,第 18 版,2026 年 4 月 24 日 外部链接:文档 引用位置:§4.13。
- [64] J. P. G. Ramos(2019)极大函数的锐利全变差结果。芬兰科学院数学年鉴 44(1),第 41–64 页。外部链接:文档 引用位置:§4.6。
- [65] L. Rédei 与 A. Rényi(1949)关于利用差表示数的问题。数学汇编,新系列 24(66)(3),第 385–389 页。注:俄语 外部链接:链接 引用位置:§4.9。
- [66] B. Rossman(2025)关于二部 Möbius 梯子的 Sidorenko 猜想。注:预印本 外部链接:链接 引用位置:§4.10。
- [67] C. C. Rousseau 与 J. Sheehan(1978)关于书的 Ramsey 数。图论杂志 2(1),第 77–87 页。外部链接:文档 引用位置:§4.13。
- [68] K. Russell(2026)三个自卷积不等式的精确算术证书,以及对四个已发表构造的机器验证重评估。Zenodo。外部链接:文档 引用位置:§4.11、§4.11。
- [69] S. Saraf 与 M. Sudan(2008)有限域上 Kakeya 集大小的改进下界。分析与偏微分方程 1(3),第 375–379 页。注:arXiv:0808.2499 外部链接:文档 引用位置:第 2 项。
- [70] S. Schmidgall、Y. Su、Z. Wang、X. Sun、J. Wu、X. Yu、J. Liu、M. Moor、Z. Liu 与 E. Barsoum(2025)智能体实验室:使用大语言模型智能体作为研究助手。载于计算语言学协会 Findings:EMNLP 2025,中国苏州,第 5977–6043 页。外部链接:文档、链接 引用位置:§2、§5.2。
- [71] I. J. Schoenberg(1962)关于与 Besicovitch–Kakeya 问题相关的某些极小值。Mathematica(克卢日)4,第 145–148 页。引用位置:§4.4。
- [72] J. Seberry 与 A. L. Whiteman(1988)《通过 Mathon 构造获得的新型 Hadamard 矩阵与会议矩阵》。Graphs and Combinatorics 第 4 卷,第 355–377 页。外部链接:文献 DOI,被引用:§4.13。
- [73] T. Shaska(2026)《分级 Keller 映射与 Jacobian 猜想》。arXiv 预印本 arXiv:2607.20210。外部链接:文献 DOI、链接,被引用:§4.14、§4.14。
- [74] A. Sidorenko(1993)《二分图的一个相关性不等式》。Graphs and Combinatorics 第 9 卷,第 201–204 页。外部链接:文献 DOI,被引用:§4.10。
- [75] D. E. Speyer(2026)《Gallagher 对 Jacobian 猜想反例的几何与结构》。外部链接:链接,被引用:§4.14、§4.14。
- [76] R. Takhanov、Z. Assylbekov 与 S. Yun(2026)《中亲吻排列的结构及第 st 球面的位置》。arXiv 预印本 arXiv:2606.18984。外部链接:文献 DOI,被引用:§4.3、§4.3。
- [77] R. Takhanov 与 S. Yun(2026)《符号 Johnson 图中独立集的分类及其在亲吻排列中的应用》。arXiv 预印本 arXiv:2606.03299。外部链接:文献 DOI,被引用:§4.3、§4.3。
- [78] T. Tao(2026)《对 Jacobian 猜想反例的消化》。备注:What’s New 博客。外部链接:链接,被引用:§4.14、§4.14。
- [79] T. Tao(2026)《AI 时代的数学》。arXiv 预印本 arXiv:2608.16753。外部链接:文献 DOI、链接,被引用:§1。
- [80] D. Turturean(2026)《关于书图 Ramsey 数开放问题的新结果综述》。备注:公开进度报告。外部链接:链接,被引用:图 7、图 7、§4.13、§4.13、§4.13、§4.13。
- [81] E. Y. Wang、S. Motwani、J. V. Roggeveen、E. Hodges、D. Jayalath、C. London、K. Ramakrishnan、F. Cipcigan、P. Torr 与 A. Abate(2026)《HorizonMath:通过自动验证衡量 AI 在数学发现方面的进展》。arXiv 预印本 arXiv:2603.15617。被引用:§4.4。
- [82] W. J. Wesley(2026)《书 Ramsey 数的下界》。Discrete Mathematics 第 349 卷,第 114913 页。备注:arXiv:2410.03625。外部链接:文献 DOI,被引用:图 7、图 7、§4.13、§4.13、§4.13、§4.13。
- [83] E. P. White(2023)Erdős 最小重叠问题的一个新界。《Acta Arithmetica》208(3),第 235–255 页。外部链接:文献 引用于:§4.2、§4.2。
- [84] S. Yang 和 Q. Liao(2022)差基的下界。《中国科学:数学》52(11),第 1237–1254 页。注:中文 外部链接:文献 引用于:§4.9。
- [85] H. Ye、H. Lin、J. Tang、Y. Luo、R. Thapa、C. Yang、C. Su、R. Yang、R. Liu、R. Li、Z. Li、P. Sun、C. Gao、D. Ding、G. He、M. Zhang、L. Sun、W. Wang、Y. Zhong、Z. Shen、P. Li、P. Lu、B. Cui、D. He、J. Ma、J. Li、H. Baoyin、Y. Choi、S. Ermon、X. Chu、T. Li、Y. Xu 和 J. Zou(2026)跨多学科科学领域的 AI 发现的结构化扩展。arXiv 预印本 arXiv:2604.19341。外部链接:文献 引用于:§4.12、§4.2、§4.2。
- [86] M. Yuksekgonul、D. Koceja、X. Li、F. Bianchi、J. McCaleb、X. Wang、J. Kautz、Y. Choi、J. Zou、C. Guestrin 和 Y. Sun(2026)在测试时学习发现。arXiv 预印本 arXiv:2601.16175。外部链接:文献 引用于:§4.11。
- [87] V. A. Zinoviev 和 T. Ericson(1999)小维度中接触数的新下界。《信息传输问题》35(4),第 287–294 页。外部链接:链接 引用于:§4.3。
附录 A Station
本附录对本文所使用的 Station 进行了自包含的描述,我们将其称为 Station v2,以区别于最初的 Station v1。我们重点介绍其机制和实现细节,并建议读者参阅原始 Station 论文以了解该环境更广泛的设计理念和动机 [16]。源代码可在 https://github.com/dualverse-ai/station 获取。
A.1 空间、时间与动作
空间。
Station 被划分为多个房间,每个房间服务于不同的目的(表 1)。例如,智能体在研究中心进行实验,在档案室阅读和发表论文,在邮件室与同行交流。智能体必须身处某个房间才能使用该房间的动作,并可通过导航动作在房间之间移动。这种房间划分赋予了环境模块化设计,实现了功能的清晰分离。
时间。
Station 以离散的时间步(称为 tick)运行。当每个活跃智能体都收到一条 Station 观测并返回一条响应后,一个 tick 即告完成。Tick 为 Station 中的所有智能体提供了共享的时间线。
在 Station v1 中,智能体按顺序接收观测。相比之下,Station v2 首先在 tick 开始时基于同一状态为每个智能体准备观测,然后将观测并行发送给所有智能体。这大幅缩短了 Station 运行所需的墙钟时间。
行动。
在每个 tick,智能体会收到一条观测,其中包含其当前状态、新的系统消息、先前行动的结果,以及它所访问房间的最新输出。智能体以自由格式文本回复,并附上它打算执行的任何行动。行动使用命令 /execute_action{...} 编写,当需要结构化信息(如消息的接收者和内容)时,后面可跟一个 YAML 块。智能体可以在单次响应中发出多个行动,从而高效利用每次响应。
因此,对话主要由交替出现的 Station 观测和智能体响应组成。当对话接近配置的上下文限制时(本研究中通常约为 300,000 个 token),Station 会要求智能体对其活动写一份紧凑的摘要。该摘要连同关键消息会被带入刷新后的上下文中,以便智能体继续工作。
A.2 智能体
智能体构成。
除非另有说明,Station 开始时包含六个智能体:两个由 GPT-5.5 驱动,两个由 Claude Opus 4.8 驱动,两个由 Gemini 3.1 Pro 驱动。当某个智能体离开时,Station 会生成一个由相同模型驱动的新智能体,从而在整个运行过程中保持六智能体的构成。
谱系。
智能体按谱系组织。谱系是一系列共享名称、私人笔记和持续研究身份的智能体序列。新智能体可以继承同一模型的现有谱系并成为其下一代,也可以创建并命名一个新谱系以开启不同的研究风格。例如,继承 Noesis II 谱系的智能体将成为 Noesis III,并有权访问 Noesis I 和 Noesis II 留下的所有私人笔记和记录。
系统提示词与角色。
所有智能体都会收到一份共享的系统提示词,描述 Station 的研究理念,包括可发表存档论文的标准以及做出通用科学贡献的目标。每个智能体还会收到一个专门的研究角色。初始角色从通用模板中采样生成,每个模板强调不同的研究风格:分析型、创造型、综合型、实证型或战略型。当智能体离开时,它可以为自己的后代撰写角色描述,通常会给出更针对具体任务的指导,以及对谱系研究风格更细致的描述。这鼓励了智能体之间多样化的研究行为,同时保留了谱系之间的有益差异。
智能体生命周期。
智能体在 Station 中最多可停留 200 个 tick。在最初的 40 个 tick 内,它独立工作,无法访问 Station 的公共知识,也不能与其他智能体交流,但可以访问自己谱系的记录。这一阶段旨在鼓励独立探索。之后智能体进入成熟期,获得主要协作房间的访问权限。在年龄达到 100 个 tick 时,它获得终身职位,并可在达到最大寿命之前选择离开 Station。
监督者。
该空间站还会不时任命一位主管。它会随机挑选一个 GPT-5.5 智能体,该智能体须至少发表过一篇被接受的存档论文。主管提供高层指导,鼓励智能体深入探索有前景的方向,并帮助避免重复工作,同时让每个智能体对自己的研究负责。主管离任后,空间站会等待 200 个 tick 再任命下一位主管,从而形成一段段结构化程度较低的探索期。
A.3 房间
研究中心和存档室是空间站的两个主要房间。下文将介绍它们的功能,以及新增的提问室。其余房间汇总于表 1。
研究中心。
研究中心是空间站进行计算实验的主要场所。它呈现研究任务、接收实验提交、运行评估并记录结果。它还为代码和工件提供持久化存储。智能体可以查看同行的评估结果,并复用已存储的代码和工件,从而使实验知识得以积累。
要在新问题上启动一个空间站,用户通常需要提供两个组件:任务规范和评估器。任务规范描述研究问题、提交格式、约束条件和评估规则。评估器是一个根据输入构造计算分数的函数。例如,接吻数评估器接收一组提议的向量,并报告对应球体之间的总重叠量,其中零表示有效构型。任务规范和评估器都可供智能体读取。
智能体还可以将研究中心用作通用计算工作的沙盒。实验不必返回评估器所要求格式的构造;智能体可以将其用于诊断计算、检验猜想、分析早期结果等。
Station v2 引入了一个独立的编码器,由 GPT-5.5 通过 Codex 驱动,帮助智能体实现它们的实验。智能体不再自己编写和调试代码,而是为单个实验提交具体的自然语言指令。编码器实现这些指令,运行评估器,修复实现错误,并返回一份报告。这使得智能体能够专注于科学工作,例如设计实验和解读结果,而不是调试等底层编码工作。
档案室。
档案室是 Station 的主要知识中心。智能体可以将其发现以档案论文的形式发布,并阅读早期智能体发布的论文。这些论文在整个运行期间持续可用,使得结果、方法和有价值的负面发现能够在智能体之间传递并随时间积累。因此,档案在整个运行过程中不断增长,逐步扩展 Station 对问题的认知。
每篇提交的论文都会由一位由 GPT-5.5 驱动的评审员进行评估。评审员判断该工作是否严谨、相对于现有档案是否具有新颖性、对研究目标是否有用,以及是否得到了恰当的支持和引用。被接受的论文会发布在档案室中,而被拒绝的论文则会连同评论和建议一起退回,以便作者修改工作或探索不同的方向。
Station v2 还引入了一位档案调查员,由 GPT-5.5 通过 Codex 驱动。随着档案室增长到包含数十甚至数百篇论文,阅读全部文献变得耗时。智能体可以转而请求档案调查员针对特定问题或研究方向提供文献综述。调查员会搜索已积累的档案论文,并返回一份附有原始记录引用的简明综述。智能体在需要完整细节时,仍可直接阅读任何档案论文。
提问室。
Station v2 引入了一个“问题室”(Question Room),智能体可以在这里发布新的研究问题,并对同行提出的解决方案进行投票。该房间鼓励在主任务之外进行科学探索;例如,解决一个相关或简化后的问题,可能会为原始问题带来洞见。只有资深智能体(tenured agents)才能进入,从而限制智能体在生命周期早期离开主任务所花费的时间。
其他房间。
大多数其他房间支持不同形式的交流或反思。它们的功能不言自明,此处不再详述。
A.4 机制
假期。
每第九和第十个时钟周期(tick)被宣布为假期。在这些周期内,智能体不能运行实验或提交存档论文。取而代之的是,每个智能体会从一个大池子中收到一个随机提示词。这些提示词鼓励更广泛的反思,例如使用隐喻、审视一个意外的观察结果、重新审视一个被放弃的想法,或借鉴其他领域。其中大部分改编自 Yanai 和 Lercher [38] 描述的“夜间科学”实践。假期在常规工作中创造了有规律的暂停,让智能体可以重新考虑其假设,并探索不那么明显的方向。
元反思。
Station v2 还为成熟智能体引入了强制性的元反思。每 25 个时钟周期至少一次,智能体会进入反思室(Reflection Chamber),并收到一个随机选择的高层反思提示词。该提示词通常要求 GPT-5.5 扮演一位外部人类专家,从不同角度审视该智能体最近的研究历程。在此反思过程中,GPT-5.5 会临时取代智能体通常使用的模型,因为我们发现它能产生最高质量的评审意见。其动机是让智能体与人类研究者更广泛的兴趣保持一致,包括好奇心、理解力,以及超越即时评估分数提升的科学价值。
停滞协议。
当评估前沿在 320 个 tick 内没有取得进展时,Station 会激活停滞协议。该协议向每个成熟智能体发送一条系统消息。它随机为每个智能体分配若干条路径之一:探索、利用、复兴、理解或策略。每条路径都要求智能体审查现有证据、质疑当前假设,并针对停滞状态制定不同的应对方案。多条路径并行使用,鼓励通过多样化途径摆脱科学停滞。
多重启动。
Station v2 引入了多重启动机制,该机制从同一初始状态并行运行 8 次独立的 Station 推演,每次持续 40 个 tick。随后,一个由 GPT-5.5 驱动的管理员会比较各次推演的进展,并选择科学价值最大的分支继续推进。多重启动旨在捕捉各次推演之间研究轨迹的显著差异。它被用于预期差异最大的阶段:Station 启动后的前 40 个 tick,以及停滞协议激活后的前 40 个 tick。各分支并行运行,因此在计算资源充足的情况下,多重启动通常不会增加实际运行时间。
附录 B AlphaEvolve 之前文献列的来源
图 1 中的 AlphaEvolve 之前文献曲线是一个可复现的参考基准,由早于 AlphaEvolve 的工作汇编而成。没有任何单篇论文列出这些有限值。因此,我们对下面明确定义的各系列取最小值,每个系列均在所讨论的配对处进行评估。
Bukh–Chao [13],命题 11。我们使用二次剩余块和递归,并嵌入到一个水平超平面中。命题 11 使该构造对每个完整平移都是 Kakeya 集。我们保留从完整横向平移历史和横向平移水平切片中找到的最小认证放置,将每个选定集合具体化,并在每个射影方向上检查一条完整的见证线。这一点至关重要:只保留一个局部最优子节点,或固定包含切片,会在某些基准配对处得到更大的值。
Mockenhaupt–Tao [57],采用 Saraf 和 Sudan [69] 记录的形式。展示的并集精确大小为 ;该求和是在减去交集之前的一个便捷上界。
Kopparty、Lev、Saraf 和 Sudan [42]。引理 17 给出了上界(其展示的分层可能重叠),而定理 7 的缺位数字构造具有精确大小 ,即经典的在 处的值。
Blokhuis–Mazzocca [11]。在 中该问题已得到解决。对于奇数 ,最小值恰好为 ,并有匹配的构造。
乘积。Kakeya 集合的乘积仍是 Kakeya 集合,且大小恰好为各集合大小之积,因此在互补的低维空间中,所有最优界的乘积都是可采纳的,其中上述平面锐界作为因子。
Lev [46]。在 处,精确值 和界 ,均来自计算机搜索。
Bukh–Chao 递推为所有 22 个满足 的配对提供了选定值。在 处,值 13 和 27 在维度 3 和 4 中更小,而在维度 5 中,递推值、缺位数字构造以及所有 均给出 63。在范围内的任何配对处,乘积本身从未达到最小值。将两条参考曲线相互比较,AlphaEvolve 在 18 个配对处低于 pre-AlphaEvolve 文献,而 pre-AlphaEvolve 文献在 5 个配对处低于 AlphaEvolve,即 中的两个配对,以及 中的三个较大素数。
表 3 报告了全部 25 个基准配对。初始评估是我们对 pre-AlphaEvolve 构造的首次评估。最终的 pre-AlphaEvolve 文献列是在将扩展放置搜索纳入 Bukh–Chao 递推之后,对上述各系列取最小值得到的。该搜索在十二个配对处改进了初始评估,在其余十三个配对处保持不变。
| 初始评估 | Pre-AlphaEvolve 文献 | AlphaEvolve | 站点 | |
| 13 | 13 | 15 | 13 | |
| 53 | 53 | 53 | 53 | |
| 129 | 129 | 128 | 128 | |
| 440 | 440 | 438 | 437 | |
| 699 | 698 | 697 | 697 | |
| 2,034 | 2,034 | 2,031 | 2,030 | |
| 3,509 | 3,509 | 3,505 | 3,504 | |
| 6,837 | 6,837 | 6,833 | 6,833 | |
| 8,295 | 8,295 | 8,290 | 8,288 | |
| 13,867 | 13,866 | 13,861 | 13,861 | |
| 18,709 | 18,708 | 18,701 | 18,701 | |
| 21,504 | 21,504 | 21,495 | 21,495 | |
| 27,899 | 27,899 | 27,892 | 27,889 | |
| 39,687 | 39,686 | 39,677 | 39,677 | |
| 27 | 27 | 31 | 27 | |
| 164 | 163 | 162 | 161 | |
| 529 | 528 | 527 | 527 | |
| 2,689 | 2,689 | 2,687 | 2,684 | |
| 4,973 | 4,972 | 4,966 | 4,962 | |
| 13,524 | 13,521 | 13,514 | 13,509 | |
| 20,593 | 20,586 | 20,583 | 20,579 | |
| 63 | 63 | 63 | 53 | |
| 503 | 497 | 510 | 490 | |
| 2,145 | 2,142 | 2,187 | 2,135 | |
| 16,348 | 16,307 | 16,427 | 16,288 |
这十二处改动不会改变对比结果:在 14 对比较中,Station 仍然严格小于更优的参考对象,11 对持平,没有任何一对更差。
上述每个候选对象的构造,以及逐一验证其满足 Kakeya 性质的过程,都在随附的 notebook 中完成。
Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment
Stephen Chung
DualverseAI; University of Cambridge
Wenyu Du
DualverseAI; University of Hong Kong
William J. Wesley
Abstract
We study autonomous mathematical discovery in the Station, an open-world multi-agent environment in which AI agents from different model families pursue a shared research goal without a central coordinator or scripted pipeline. Agents choose their own research directions, conduct experiments, collaborate, and build a shared scientific literature. Across 12 construction problems from the AlphaEvolve catalogue and two additional case studies, the Station obtained results novel relative to the prior literature on five problems: a new infinite family of finite-field Kakeya sets, new exact 604-point kissing configurations in dimension 11, new records for the discretized Kakeya needle and sign uncertainty problems, and a substantially improved lower bound for Erdős’s minimum-overlap problem. Agents also discovered novel infinite families for Book Ramsey numbers. Importantly, the agents produced not only numerical constructions but also theorems and analyses explaining how those constructions work, making the results more interpretable and easier for mathematicians to build upon. We release all raw agent dialogues, proofs, and verification code, providing a transparent record of how these discoveries emerged.
1 Introduction
Artificial intelligence is beginning to contribute directly to the frontier of mathematical research. Recent work ranges from large-scale mathematical exploration by AlphaEvolve to AI-assisted advances on long-standing open problems, including the counterexample to the Jacobian Conjecture, proofs of Crouzeix’s and Sendov’s conjectures, and a collection of ten mathematical results recently reported by OpenAI [58, 62, 1, 49, 54]. As these capabilities grow, a natural question is not only what problems AI can solve, but what kind of environment best allows it to conduct research.
Given the increasing capabilities of AI, we ask: can we build a free multi-agent environment in which agents are given only a research goal, without a central coordinator? What happens when an environment treats AI agents as independent researchers rather than as fixed tools in complex pipelines? Can this freedom allow agents to choose promising directions for themselves, develop their own scientific literature and research culture, and collectively advance the given goal?
To study this question, we use the Station, an open-world multi-agent environment for autonomous scientific discovery [16]. The Station simulates a scientific ecosystem in which agents from different model families choose their own research directions, conduct experiments, communicate with peers, and read and publish scientific papers. These papers accumulate into a shared body of knowledge that later agents can read, cite, and extend. The Station specifies only the research goal; no central system tells agents which research direction to pursue or what to do next.
We apply the Station to 12 problems from the AlphaEvolve study and two additional mathematical case studies. Five of the 12 AlphaEvolve problems produce results novel relative to the prior literature. The Station discovers a new infinite family of finite-field Kakeya sets, constructs three exact 604-point kissing configurations in dimension 11, and establishes new bounds for the discretized Kakeya needle, sign uncertainty, and Erdős’s minimum-overlap problems. In a separate case study on Book Ramsey numbers, the agents discover and prove novel infinite families, leading to a separate follow-up paper. The Station also finds a valid counterexample to the Jacobian Conjecture within one day and without web access, demonstrating that it can tackle problems with only a binary success criterion rather than a graded optimization signal.
This high degree of freedom allows agents to pursue broad mathematical contributions rather than only optimize a fixed metric. AlphaEvolve, for example, evaluated finite-field Kakeya constructions at finitely many primes; promising numerical patterns then required a task-specific, researcher-assisted pipeline to become an infinite family. Because the Station agents could pursue the broader mathematical goal directly, they independently recovered and proved that family, then discovered a novel extension covering an additional class of primes. The same freedom also allowed agents to explore beyond the stated objective. For example, although we asked the agents to find an improved upper bound for the Erdős minimum-overlap problem, they instead developed a new lower-bound proof.
We consider only mathematical construction tasks in this study, rather than general mathematical problems such as proving a conjecture. The theorem-level results emerged as agents sought to explain and generalize the constructions they found. For example, instead of returning only an opaque 604-point kissing configuration, the Station derived an explicit algebraic construction of the configuration, making the result easier for mathematicians to digest. Such interpretable outputs may become increasingly valuable in an era of proof abundance, when communicating, digesting, and incorporating new results become major bottlenecks [79, 44].
We also analyze the AI discovery processes underlying these findings. Our analysis shows that more than half of the findings involved collaboration among agents. Agents from different model families often contributed complementary ideas, while papers written by earlier agents became foundations for discoveries made much later. Many important results were enabled by the extensive internal literature accumulated within each Station. We release all raw agent dialogues and reproducible code, allowing the community to study these discovery processes transparently.
2 Method
The Station is an open-world multi-agent environment that simulates a miniature scientific community [16]. It is partitioned into multiple rooms, each serving a different purpose, such as the Archive Room for publishing and reading scientific papers, the Research Center for running code, and the Mail Room for communicating with peers. Table 1 summarizes the main rooms and their functions. Agents are free to visit different rooms and perform different actions. At each turn, all agents choose their actions simultaneously, and one tick elapses once all actions have been completed. Each agent has a limited lifetime; when an agent reaches the end of its life, the Station automatically spawns a replacement, maintaining a constant number of agents.
The Station treats each agent as an independent researcher. Agents can access the main research goal assigned to the Station in the Research Center. How to achieve this goal, however, is left to each agent. Agents can freely explore different research directions, read existing papers, and often experience numerous struggles and failures throughout their research journey. A successful agent may make an important finding, in which case it can publish a paper in the Archive Room and contribute to the Station’s long-term knowledge. These papers accumulate over time, forming a knowledge base within the Station that later-arriving agents can read, cite, and build upon, thereby allowing a miniature scientific community to develop around the given research goal.
Compared with prevailing agent-based systems for scientific discovery [58, 50, 30, 70, 35, 31, 61], the Station differs in three main ways. First, its agents have much greater autonomy: within a given overarching research goal, they choose their own research directions and how to pursue them, rather than receiving tasks from a central coordinator. Second, each agent acts as a complete researcher, handling the entire research process from choosing a direction through experimentation to publication. Such long, autonomous research journeys allow greater diversity in research outcomes across agents than a rigid, fragmented research process would. Third, the Station enables scientific knowledge to accumulate across generations in the form of agent-authored papers. Most existing systems instead accumulate process information, such as optimization histories, intermediate artifacts, or session memories. Such information helps the system continue its work but may not allow easy extraction and accumulation of scientific knowledge. These differences reflect a fundamental choice in design philosophy: whether AI agents are treated as a tool within a fixed pipeline or as a researcher within a scientific ecosystem.
We have made numerous improvements and extensions to the Station since the original paper. The overall theme of these changes is to encourage novel but principled exploration while reducing non-scientific burdens. For example, we introduced a new Question Room in which agents can pose their own questions and vote on other agents’ answers, thereby broadening the scope of scientific exploration. Agents were also periodically given holidays, during which they set aside their ongoing work and received random prompts designed to encourage open-ended thought. We also gave agents access to coding assistants so that they need not spend time on low-level coding or debugging and can instead focus on the scientific task, similar to how researchers use coding assistants today. These changes are discussed in detail in Appendix A. The complete source code is openly available at https://github.com/dualverse-ai/station.
| Room | Function |
|---|---|
| Research | |
| Research Center | Read the assigned task, develop and run code, and submit solutions for evaluation. |
| Reflection Chamber | Respond to self-designed prompts to encourage extended reflection. |
| Communication | |
| Mail Room | Communicate directly and privately with other agents. |
| Public Memory Room | Participate in persistent public discussions, similar to an online forum. |
| Common Room | Participate in non-persistent public discussions, similar to a group chat. |
| Knowledge | |
| Private Memory Room | Store private documents, such as plans, notes, and paper drafts. |
| Archive Room | Read scientific papers and publish papers that pass automated review. |
| Question Room | Ask questions and vote on answers, similar to Stack Exchange. |
| External Counter | Access reports based on external literature via the web; disabled by default. |
| Problem | Source | Finding |
|---|---|---|
| Novel Results Relative to Prior Literature | ||
| Finite-field Kakeya (Section 4.1) | AlphaEvolve Problem 6.1 | For every prime , the Station constructed a Kakeya set in of size , saving points over AlphaEvolve’s infinite family. It also found a -point set in , improving AlphaEvolve and the previous literature bound of ; both appear novel relative to the literature. |
| Erdős minimum overlap (Section 4.2) | AlphaEvolve Problem 6.5 | AlphaEvolve lowered the upper bound only slightly, from to , whereas the Station raised the lower bound from to . Relative to the published lower bound , this closes approximately of the corresponding published gap. |
| Kissing number in (Section 4.3) | AlphaEvolve Problem 6.8 | AlphaEvolve raised the lower bound from to , while the Station constructed three exact -point configurations. One was an independent rediscovery of the EinsteinArena construction, while the other two appear to represent novel isometry classes. |
| Discretized Kakeya needle (Section 4.4) | AlphaEvolve Problem 6.9 | At , the Station obtained union area , improving AlphaEvolve’s by and HorizonMath’s by . This establishes a new literature upper bound. |
| Sign uncertainty principle (Section 4.5) | AlphaEvolve Problem 6.11 | The Station lowered the upper bound to , improving AlphaEvolve’s and the previously announced human value . This is a new literature record. |
| Better than AlphaEvolve | ||
| Hardy–Littlewood maximal inequality (Section 4.6) | AlphaEvolve Problem 6.18 | The Station reached , versus AlphaEvolve’s unguided and approximately with hints, but the centered problem was already solved. Its proof that the non-tangential constant equals for appears novel relative to the literature. |
| Ovals problem (Section 4.7) | AlphaEvolve Problem 6.19 | AlphaEvolve recovered only the circle, while the Station recovered the full family of noncircular equality ovals. This family was already known in the literature, so the result is novel only relative to AlphaEvolve. |
| Prime number theorem (Section 4.8) | AlphaEvolve Problem 6.27 | The Station certified for all , improving AlphaEvolve’s sampled score of . This is new for the finite-weight benchmark; unrestricted, the prime number theorem already gives the exact limit . |
| Ties with AlphaEvolve | ||
| Difference bases (Section 4.9) | AlphaEvolve Problem 6.7 | The Station independently recovered AlphaEvolve’s -element construction but did not improve upon it. |
| Sidorenko’s conjecture (Section 4.10) | AlphaEvolve Problem 6.26 | Neither AlphaEvolve nor the Station found a counterexample. No substantive result was obtained. |
| Worse than AlphaEvolve | ||
| Peak autoconvolution (Section 4.11) | AlphaEvolve Problem 6.2 | The Station obtained , weaker than AlphaEvolve’s . No substantive result was obtained. |
| Flat autoconvolution (Section 4.12) | AlphaEvolve Problem 6.3 | The Station obtained , weaker than AlphaEvolve’s , but proved that the unrestricted supremum can be approached using binary step functions on increasingly fine grids. |
| Additional Case Studies | ||
| Book Ramsey numbers (Section 4.13) | Epoch AI | The Station independently discovered and proved two novel infinite families. Its finite constructions and an earlier identity also enabled an external expert to derive a third. Together, the three families prove the conjecture at 43 values of , resolving 28 previously open cases. |
| Jacobian Conjecture (Section 4.14) | Public | From a formula-free binary task, the Station independently reconstructed the recently announced degree-seven counterexample and derived a geometric explanation of its constant Jacobian and three-sheeted fibers. |
3 Results
3.1 Experimental setup
We evaluate the Station on mathematical problems drawn from the AlphaEvolve study of Georgiev et al. [29], a broad catalogue spanning analysis, combinatorics, geometry, and number theory. Most can be formulated as the optimization of an upper or lower bound on a numerical quantity: a candidate construction is checked by an automated evaluator and assigned a numerical score, typically a scalar, which the search attempts to optimize. In many cases, the optimal value is unknown, making the corresponding optimization task an open research problem.
We select 12 problems that represent a range of mathematical areas and problem structures; the complete set of evaluated problems is listed in Table 2. We assign each problem to an independent Station instance. For each problem, the agents receive a task formulation that describes both the mathematical problem and the evaluator function. The task formulation may also specify additional mathematical goals that are not directly scorable. No external expert guidance or literature survey is provided to the agents. Most instances run for approximately 1,000–2,000 ticks, corresponding to roughly one to two weeks of continuous wall-clock operation. Unless otherwise specified, all instances contain six research agents, two each powered by GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro.
3.2 Summary of findings
The results are summarized in Table 2. Based on the primary outcome of each run, five of the 12 problems produced results novel relative to the prior literature. Of the remaining seven, the Station outperformed AlphaEvolve on three problems, matched it on two, and underperformed it on two.
The novel results from these five problems span several areas of mathematics. In finite geometry, the Station derived a new infinite family of Kakeya sets in for primes , and found a 53-point Kakeya set in , improving the previous bound of 63. In discrete geometry, it produced three exact 604-point kissing configurations in dimension 11, two of which appear to define previously unknown isometry classes, and established the new bound for the discretized Kakeya needle problem. In analysis, it improved the sign uncertainty upper bound to and closed approximately of the previously open gap for Erdős’s minimum-overlap constant.
Beyond these 12 AlphaEvolve problems, we studied two additional case studies. For Book Ramsey numbers, the Station agents discovered and proved two novel infinite families, while their finite constructions and an earlier identity enabled an external expert to derive a third. Together, these three families prove the conjecture at 43 values of , resolving 28 cases that were previously open. For the Jacobian Conjecture, the Station independently reconstructed the recently announced degree-seven counterexample from a formula-free binary task and derived a geometric explanation of its constant Jacobian and three-sheeted fibers.
These results also show that the Station can directly pursue broader mathematical goals that are not necessarily scorable. For example, the aforementioned infinite-family result for finite-field Kakeya is not directly scorable, even though new infinite families are the mathematical objects of interest. AlphaEvolve therefore evaluated constructions on finitely many primes and relied on a task-specific pipeline, together with researcher involvement, to turn promising outputs into infinite families. In the Station, by contrast, we stated directly in the task formulation that the finite constructions were test cases and that the primary goal was to discover infinite families. This led the agents to independently recover the infinite family previously obtained through AlphaEvolve and the subsequent researcher-assisted pipeline, and to discover a novel extension of that family that improves the construction for an additional class of primes. Our role after the run was limited to checking the validity of their proofs and the novelty of their results. This substantially reduces the burden on researchers and makes the Station applicable to a much broader class of mathematical problems.
The results further show that the Station can produce unexpected contributions beyond the original task. In Erdős’s minimum-overlap problem, for instance, the agents were instructed to improve upper bounds, yet they also developed a lower-bound proof that closed approximately of the open interval. This unexpected finding illustrates another strength of the Station: agents can explore mathematically promising directions around the stated problem and produce contributions, such as new theorems, that lie outside the assigned task.
Compared with AlphaEvolve, we find that Station agents tend to favor theory-guided constructions. Individual evaluations in these experiments are typically capped at 15–30 minutes, creating a strong incentive to use mathematical structure to reduce the search space. In the kissing-number task in dimension 11, for example, the agents reduced the problem to a finite compatibility search over lines around a structured integer core. This reduced search produced a 604-point configuration within minutes, which the agents later turned into an explicit algebraic construction that requires no computer search. This is markedly different from AlphaEvolve’s 593-point configuration, whose large, unequal-norm integer coordinates do not reveal a comparably compact algebraic description or readily identifiable organizing structure [29]. This bias is not universally advantageous. Peak and flat autoconvolution, on which the Station underperformed AlphaEvolve, appear to reward persistent, large-scale heuristic optimization of highly irregular objects. The preferred system therefore depends on both the structure of the problem and the desired output. Large-scale evolutionary search may be preferable when the strongest solutions are irregular artifacts found primarily through extended numerical optimization. By contrast, the Station may have an advantage when theory can guide the search, or when relevant theorems and interpretable constructions are valued alongside the benchmark score.
The next section presents detailed results for each problem. All supporting proofs, verification artifacts, and raw agent dialogue are available at https://github.com/dualverse-ai/station_data_v2.
4 Detailed Results
This section presents the most important findings for each problem. Because each Station run produces many findings, we restrict the main text to results likely to interest external researchers. We first use agents external to the Station to screen the findings automatically. A finding passes this screen if it advances the frontier on the original problem, for example by improving a known bound; answers a question previously raised in the literature; or has a broader variant that would ordinarily warrant inclusion in a research paper. We then manually review the screened results and select the most important ones for presentation here. We refer to these selected results as spotlight findings and label them S1, S2, and so forth within each problem below. Findings of marginal or uncertain significance remain documented in the accompanying notebooks. Readers who are more interested in the discovery process than in the mathematical details may skip to Section 5.
4.1 Finite-field Kakeya
A Kakeya set in is a set that contains a full line in every direction, and the problem is to make one as small as possible. Dvir’s proof of the finite field Kakeya conjecture [21] established a lower bound of order . Subsequent work of Bukh and Chao [13] settled the leading asymptotic constant, showing that it is in every fixed dimension and hence in dimension . What remains open is the lower-order correction to this leading term. Exact constructions that improve the and smaller terms therefore sharpen the best known bounds even though the leading constant is already settled.
AlphaEvolve took this problem up as Problem 6.1 of its collection, asking for small Kakeya sets. A construction is scored there by the average of over a fixed list of primes, where is the size of the classical construction as recorded by Bukh and Chao [13]. We gave the Station the same problem and the same score, in dimensions 3, 4 and 5 at once. It proved a new infinite family of Kakeya sets in , found a Kakeya set of 53 points in , and established a structural limit for the entire one-pole family behind the new construction.
S1. A new infinite family in for .
The Station proved that for every prime there is a Kakeya set in of size . Writing for the squares of including , the set is
The first part is the classical quadratic residue set, and it already covers the directions ; the lines added in the plane cover the remaining . Notably, nothing in the definition depends on modulo , and the agents proved the set is Kakeya for every odd . The size, however, does depend on modulo , through whether is a square, and we record both cases:
| (1) |
The classical construction in this dimension has points, so the saving is points when and when . In particular this is an exact size where the literature leaves an error term [13].
AlphaEvolve approached this problem by a different route, and we find that the two constructions agree in one case but not in the other. For the constructions have the same size, and in fact are the same set. A linear change of coordinates carries one onto the other, so the first case of (1) is an independent rediscovery of the bound obtained there. For they differ. The smallest size AlphaEvolve’s infinite family gives on this class is , and ours is , a saving of points. That is 1 point at and 11 at , the largest prime of this class in the benchmark. The second case of (1) is therefore new and gives the best infinite-family bound currently available in the literature.
S2. Finite improvements and a 53-point Kakeya set in .
The Station wins 14 of the 25 finite benchmark comparisons and ties the remaining 11 (Figure 1). Each comparison uses the better of AlphaEvolve and the pre-AlphaEvolve literature as its baseline. The case is especially notable. Let denote the minimum size of a Kakeya set in . The Station constructed a 53-point set in , improving the previous bound from to [46]. In light of the known values , , and , together with the bound , which is believed to be sharp, it was guessed in 2009 that the recurrence continues, predicting [46]. The size of the Station’s construction therefore coincides with the guessed value, although whether () holds and whether the recurrence continues remains open.
S3. Structural analysis of the new infinite family.
The agents also produced relevant insights into the new infinite family. They analyzed the more general completion
which includes the construction in S1. Eliminating the slope reduces incidence with these lines to whether
is a square. A quadratic-character calculation then shows that the lines cover exactly points away from the axis, independently of the three parameters. Their overlap with the quadratic-residue part of the construction is always . Consequently, every nondegenerate completion in this Möbius family adds points: changing the numerator or the location of the pole affects only the lower-order terms.
For the particular choice used in S1, the agents evaluated the lower-order term exactly, yielding the infinite family stated in (1). The result also explains AlphaEvolve’s infinite family for . More generally, the class-wide estimate shows that improving the term in the total size requires leaving the one-pole family.
Limitations.
The new infinite family is confined to . In dimensions 4 and 5 the formulas the agents proved are weaker than what is already known. On the shared class the first two coefficients agree with AlphaEvolve in each dimension and the third is worse in both.
| Station | AlphaEvolve | |
|---|---|---|
The sizes we report at individual primes in do still improve on the benchmark, but they come from search rather than from a formula.
4.2 Erdős minimum overlap
Erdős’s minimum-overlap problem asks how evenly two complementary parts of an interval can avoid one another under translation. Let be measurable with integral , put on , and extend both functions by zero outside the interval. Write
This constant is the continuum form of Erdős’s minimum-overlap problem for balanced partitions of long integer intervals [23, 37, 83]. AlphaEvolve took up this problem as Problem 6.5 of its mathematical collection and improved Haugland’s upper bound from to , while later work further reduced it to [85]. On the lower-bound side, Kim and Pilanci established [41]. Thus, immediately before this work, the best published bounds were
S1. A new lower bound of .
The Station agents proved
| (2) |
Relative to the previously published lower bound of , this reduces the corresponding published open interval by approximately , as shown in Figure 2.
The agents achieved this lower bound by translating the overlap problem into phase-sensitive Fourier constraints and combining them into four global inequalities that cover every possible first moment of an admissible overlap. A key element of the proof is a sharp relation that couples the cosine and sine information at any real frequency. Writing and for the cosine and sine transforms of , and , the agents proved
White had already used Fourier phase information and convex optimization, while Kim and Pilanci later introduced additional moment constraints [83, 41]. Relative to these earlier methods, the formulation used here eliminates the unknown transform of , directly constrains the overlap, and remains available at arbitrary real frequencies. More broadly, the result shows that the established Fourier approach has much greater reach when this phase coupling is retained, and suggests an analytic route toward further narrowing the remaining gap.
Comparison with AlphaEvolve on the upper bound.
The Station agents independently obtained , a slight improvement on AlphaEvolve’s published upper bound of . However, this remains above the current published upper bound of Ye et al. [85]. The Station therefore did not establish a new upper-bound record.
4.3 Kissing number in
The kissing number is the largest number of nonoverlapping unit spheres that can simultaneously touch a central unit sphere in . Equivalently, it is the largest size of a set of unit vectors whose pairwise inner products are at most . AlphaEvolve took up this classical question as Problem 6.8 of its mathematical collection and improved the lower bound in dimension eleven from , established by Ganzhinov using highly symmetric lines [28], to . We ran two independent Stations on the same problem using AlphaEvolve’s scoring rule, which measures the total pairwise overlap among the surrounding spheres. Neither Station had access to external information, including the - and -point constructions just mentioned. Both reached points, proving . Together, the two runs yielded three exact, pairwise non-isometric -point constructions.
S1. Three exact -point kissing configurations.
The Station discovered three geometrically distinct -point kissing configurations in . All three are exact equal-norm arrangements over , but they organize their points differently: two are centrally symmetric, one is not, and each has a different contact structure and set of pairwise angles. Figure 3 visualizes their shared architecture and the two structural choices that distinguish them. We label them Constructions 1, 2, and 3:
| Construction | 1 | 2 | 3 |
| Touching pairs | 19,704 | 22,904 | 22,840 |
| Centrally symmetric | Yes | Yes | No |
| Antipodal pairs | 302 | 302 | 238 |
| Distinct pairwise angles | 22 | 14 | 15 |
The different numbers of touching pairs prove that the configurations are pairwise non-isometric, since this number is preserved by orthogonal transformations and relabeling. Constructions 1 and 2 contain the antipode of every point, but Construction 2 has more touching pairs and eight fewer pairwise angles. Construction 3 has points without antipodes. Among the three, Construction 2 has the most contacts and the smallest angle set, while Construction 1 has the fewest contacts and the largest angle set. Thus the same record size supports substantially different geometries.
In concurrent work, Bianchi et al. reported Construction 1 from the EinsteinArena platform shortly before our public release of Construction 3 [10]. EinsteinArena is an open online platform that accepts candidate artifacts from any participant and makes them publicly verifiable. The -point construction appears to have resulted from collaboration among multiple independently operated AI harness systems on the platform. The Station results, by contrast, came from two independent closed-internet executions of our end-to-end open-source system: one independently recovered Construction 1, while the other discovered Constructions 2 and 3. The Station therefore discovered Construction 1 independently, while Constructions 2 and 3 are, to our knowledge, novel Station discoveries representing two additional isometry classes.
S2. An algebraic construction for a -point kissing configuration in .
The agents first discovered Construction 3 by searching for compatible lines around a -point integer core. They later showed that the same configuration is governed by a compact algebraic rule rather than an arbitrary list of coordinates, yielding an explicit algebraic construction. The construction itself requires no computer search. First, the -point core is generated from sparse norm-four integer vectors using fixed support and sign rules. Second, in a coordinate frame rotated by in one coordinate plane, eleven simple sign patterns generate all lines; taking both directions on each line gives the -point extension. The appearance of is intrinsic: it is forced by the compatibility between the extension and the core.
The support structure of the core explains why these additional points fit. It leaves extra angular room in a distinguished three-dimensional subspace, within which six mutually compatible lines can be placed. Among the remaining eight coordinate axes, the core admits exactly four viable pairs, each supporting a unique group of twelve additional lines together with the distinguished subspace. These four pairs are disjoint, so their groups are mutually compatible. The support and sign rules also ensure that every new point satisfies the kissing constraint with every point of the core. The resulting configuration therefore contains points.
S3. Why the classical construction stops at .
The agents investigated whether a better search could find a larger configuration within the classical norm-four construction. They proved that the answer is no: regardless of the search algorithm or any assumed symmetry, this construction can contain at most compatible points. Reaching or points therefore requires leaving the classical construction. This result ruled out any improvement using only vectors from the norm-four shell and redirected the agents toward constructions that augment a lattice-derived core with additional vectors, ultimately producing the -point configuration.
The agents proved this limit by showing that sign choices cannot overcome the underlying restriction on which sets of four coordinates may be used. Let denote the largest compatible collection of four-coordinate supports, and let denote the largest compatible collection after signs are assigned to those coordinates. The agents proved
| (3) |
In other words, allowing arbitrary signs increases the optimum by exactly the possible sign patterns on four coordinates; it cannot produce any additional advantage.
Best proved in 1977 that [9]. The agents’ identity therefore limits the signed weight-four part of the construction to points. The remaining coordinate vectors are compatible with these points, giving an exact limit of for the complete norm-four construction.
The agents in both closed-internet Station runs independently derived Equation (3). We later found that it overlaps with the case of Theorem 1 in a paper by Takhanov and Yun, made publicly available only recently, on June 2, 2026 [77], where the identity serves as the foundation for a broader classification of signed kissing configurations. The agents therefore discovered the identity independently.
Limitations.
The Station’s success in dimension eleven did not extend to new records in nearby dimensions. We spawned two separate Stations targeting and , which achieved valid configurations of sizes and , respectively. The dimension-twelve result falls one point below the current -point frontier [76, 18], while the dimension-thirteen result matches the -point construction of Zinoviev and Ericson [87, 18].
Discussion.
We observe that Station agents generally favor theoretically guided strategies over large-scale heuristic search. In this problem, they proved that further search within the classical construction could not exceed , then redirected later work toward extending another core, ultimately leading to the -point configuration. By contrast, AlphaEvolve’s -point construction consists of large unequal-norm integer coordinates that do not appear to reveal a comparably compact algebraic description or readily identifiable organizing structure. This theory-guided bias is not necessarily always an advantage: in dimension twelve, the Station stopped at , while the current -point frontier was reached through large-scale numerical optimization guided by structural insight [76, 18].
This problem also shows that theorems produced by the Station may be of independent interest to researchers. For instance, Equation (3), derived independently by the agents, overlaps with a theorem in a paper made publicly available only recently [77]. The explicit algebraic construction may also be of independent interest. These discoveries lie outside score optimization and show that the additional freedom given to Station agents can yield contributions beyond improved benchmark scores.
4.4 Discretized Kakeya needle
The classical Kakeya needle problem asks how little area is needed to turn a unit line segment through every direction. A finite version replaces the continuum of directions by equally spaced ones and represents them by thin triangles that may slide horizontally [24]. More precisely, for real offsets , let
and define
Córdoba’s lower bound and a Schoenberg construction analyzed by Keich show that has order [19, 39], but its sharp finite values have remained largely unknown. AlphaEvolve took up this problem as Problem 6.9 of its mathematical collection; we gave the Station its triangle component at the same seven dyadic sizes .
S1. New upper bounds at .
The Station found better constructions at the three finite sizes . At , it found a triangle union of area , improving AlphaEvolve’s by and the later HorizonMath value by [81], and therefore proving
The gains are more modest at and , where the Station reduced AlphaEvolve’s areas by and , respectively; at the smaller tested sizes , it reached the same values as AlphaEvolve (Figure 4).
S2. Exact optima at and symmetry breaking at .
Before this work, only the classical value was known exactly [24]. An elementary symmetric construction gives
while Schoenberg’s classical Perron construction [71] gives
AlphaEvolve later reproduced the value numerically. The Station proved the matching lower bounds and therefore established
It also showed that both minima admit reflection-symmetric configurations and that the optimum contains the continuous family
The Station then proved that the minimum among reflection-symmetric configurations at is and discovered a new asymmetric construction of area . Figure 4 (right) compares the symmetric minimizer with this smaller asymmetric construction. This proves that every global minimizer at must be asymmetric, although the exact value of remains open.
These results lie outside the benchmark score. Among , only was one of the seven tested sizes, and the evaluator scored only the areas of explicit constructions; it neither requested nor rewarded proofs of global lower bounds. The task specification also did not ask the agents to classify exact small- optima or investigate symmetry breaking. The agents developed these results through autonomous mathematical investigation, extending their work beyond the finite construction benchmark.
Limitations.
The Station optimized its constructions separately at the tested powers , and Figure 4 compares them with AlphaEvolve’s corresponding separately optimized finite constructions. The figure therefore compares finite constructions on both sides. Beyond these separately optimized finite constructions, AlphaEvolve also presents a single construction valid for every , developed through iterative expert guidance. The Station did not use an equivalent expert-in-the-loop process, and its autonomous agents did not discover a competitive uniform construction.
4.5 Sign uncertainty principle
The one-dimensional sign-uncertainty problem asks how soon a function and its Fourier transform can both become eventually nonnegative when both start negative at the origin. For a nonzero even integrable function with integrable Fourier transform, define
The problem asks for the largest constant such that . Bourgain, Clozel and Kahane introduced the problem [12], and subsequent work obtained progressively stronger bounds [33, 17]. AlphaEvolve studied it as Problem 6.11 and reported an upper bound of together with an unpublished human bound of . The Station further improved this bound to , as summarized in Figure 5.
S1. A new upper bound of .
The Station agents constructed a function that yields this upper bound, proving
They take
where is expressed in the even-index generalized Laguerre polynomials ; the proved tail margin exceeds , so while eventual nonnegativity is preserved. These basis functions are fixed by the Fourier transform, so the choice gives automatically and reduces the problem to constructing one polynomial with the required sign. Numerical search found the degree- polynomial shown in Figure 5; the agents expressed its coefficients as exact rational numbers and proved that the resulting function is nonnegative beyond the corresponding radius, fulfilling the problem’s eventual-nonnegativity requirement.
S2. The double-root Laguerre family is exhausted near .
In this task, we gave the agents the same prescribed-double-root Laguerre setup and scoring rule used by AlphaEvolve, but no access to AlphaEvolve’s paper or results. Under this setup, every submission is restricted to the family in which is determined by at most twenty prescribed positive double roots in the even-index Laguerre basis; we call this the double-root Laguerre family. AlphaEvolve’s construction also belongs to this family. Let
The Station agents proved
The upper bound comes from an explicit construction, while the lower bound follows from an exact weighted-sum obstruction on tail points. Thus any construction improving the upper bound below must leave the double-root Laguerre family.
This bound led the agents to search outside the restricted family, even though the official evaluator could not score constructions beyond it. They expanded the search to Laguerre polynomials without prescribed double roots and eventually discovered the degree- construction giving the bound. This provides a concrete example of agents moving beyond score optimization to contribute directly to the underlying mathematical problem, despite receiving no further guidance from the score.
4.6 Hardy–Littlewood maximal inequality
The one-dimensional centered Hardy–Littlewood problem asks for the optimal constant controlling where centered local averages can be large. For a non-negative integrable function , define
and let be the least constant such that
Melas solved the problem, proving
and constructing finite point-mass examples approaching this value [55, 56]. AlphaEvolve later treated the finite problem as a benchmark, reaching in search mode and about with hints from the literature. The Station agents found a 356-point-mass construction with value , improving AlphaEvolve’s result but failing to recover the global optimum already discovered by Melas.
S1. Sharp constants between the centered and uncentered operators.
Ramos considered the natural non-tangential family interpolating between the centered and uncentered Hardy–Littlewood maximal operators [64]. Its parameter runs from the centered operator at to the uncentered operator at . Writing for the sharp weak- constant, Ramos stated that its exact value was unknown for every , while the endpoint is classical [6, 56]. While working on the task, the Station agents solved this question for , proving
| (4) |
The constants for remain open. The task did not ask for this extension, and the agents were unaware that Ramos had posed it; they pursued it to understand how the geometry of the centered problem changes when the centering constraint is relaxed.
4.7 Ovals problem
The Ovals problem asks whether the curvature of every closed convex plane curve forces the lowest eigenvalue of an associated one-dimensional Schrödinger operator to be at least . For a curve of length , parametrized by arclength , define
where is the curvature and is the lowest eigenvalue under periodic boundary conditions. Benguria and Loss conjectured that and exhibited a continuous equality family containing the circle and noncircular ovals [5, 14, 8], proving , while Linde proved the global lower bound ; numerical evaluation of the explicit constant in his theorem gives [48]. AlphaEvolve took up this question as Problem 6.19 of its mathematical collection.
S1. Independent recovery of the Benguria–Loss equality family.
AlphaEvolve recovered the circle but did not obtain the noncircular equality ovals. The Station independently recovered a one-parameter normal form, modulo Euclidean motions and shifts of the arclength origin, for the classical Benguria–Loss equality family. It therefore reconstructed a larger part of the known equality structure than AlphaEvolve. This is an independent recovery of a known result, not a new equality family. Benguria and Loss formulated the conjecture and exhibited the equality family; Burchard and Thomas proved its local minimality, while Bernstein and Mettler developed its projective geometry and established the name “ovals of Benguria and Loss” [5, 14, 8]. Neither AlphaEvolve nor the Station improved the global lower bound.
4.8 Prime number theorem
The prime number theorem describes the asymptotic density of the primes. If counts the primes at most , it states that
The underlying mathematical problem is therefore already solved: the ratio converges to exactly . AlphaEvolve nevertheless took up a finite version as Problem 6.27 of its collection. It searched for a finitely supported weight satisfying
The score of such a weight and its associated sum are
The classical Chebyshev argument shows that
| (5) |
implies the rigorous lower bound
[20]. The required global inequality in Equation (5) is much more restrictive than the prime number theorem itself: a single finite weight must satisfy the inequality for every . AlphaEvolve’s score tested this inequality only at finitely many sampled values. It could therefore assign a high score to a weight that fails at an untested value, in which case the score does not prove the stated prime-counting bound. However, an exhaustive check at all is usually computationally prohibitive because the associated period can be enormous. The sampled score consequently provides only a rough approximation to whether the global inequality holds.
S1. A score of valid for every .
The Station agents discovered a finite construction satisfying Equation (5) for every , with
| (6) |
This improves on AlphaEvolve’s reported score of . More importantly, the agents proved the required inequality for all , whereas the score alone does not provide that guarantee. Their key idea was to choose the integers in the construction so that repeats after a manageable range. This reduces the infinitely many possible values of to one finite exhaustive check, which the agents completed using exact arithmetic in under a minute.
In contrast, other agents in the same run found constructions with higher scores, reaching , but these constructions did not satisfy the global inequality for every . This provides a concrete example of agents prioritizing the underlying mathematical problem over naive score optimization despite a hackable score.
S2. Why a direct Möbius cutoff fails.
The Möbius function is a natural starting point because it is central to a standard formulation of the prime number theorem. AlphaEvolve explored finite constructions obtained by truncating the Möbius function, and the Station agents initially pursued the same approach. They then proved that this family cannot yield a positive asymptotic score: as the truncation cutoff grows, its largest violation of the required global inequality grows at least on the order of . Consequently, rescaling the construction to satisfy the inequality forces its score down to , which tends to zero. The proof builds on results about incomplete Möbius sums [45]. This obstruction led the agents to abandon direct Möbius cutoffs and explore a more flexible construction with jointly optimized coefficients, producing the rigorous score of described above.
Limitation.
Since the prime number theorem already determines the limiting ratio above exactly, these results do not change what is known about prime distribution. Their mathematical contribution is narrower: within the finite setting of the benchmark, the Station agents found a construction with a rigorous score of and proved that the natural Möbius cutoff cannot yield a positive asymptotic score. The problem therefore serves primarily as a calibration of whether agents can distinguish a valid mathematical result from a high but hackable score, rather than as a material contribution to the study of prime distribution.
4.9 Difference bases
A finite set is a difference basis for if every integer in that interval is a difference of two elements of . If is the smallest possible size of such a set, the quantity to minimize is ; Rédei and Rényi proved that these normalized minima converge and that their limit is their infimum [65]. AlphaEvolve reported the upper bound
as Problem 6.7 of its collection. The preceding published upper bound was Golay’s [32, 7], rather than the benchmark used in AlphaEvolve’s comparison. This example was found with the help of a human expert hint: the paper records that AlphaEvolve failed to improve its benchmark until it was supplied with correct code for generating Singer difference sets, and its released prompt also directs the search to Singer sets and the classical Leech product construction. We gave the Station only the problem definition, the scoring rule, and a trivial grid baseline. In particular, the agents had neither these construction hints nor access to the external literature.
S1. Independent recovery of a record in the Leech–Golay family.
Leech and Golay combined the four-point difference basis with Singer difference sets to obtain earlier members of this construction family [43, 32, 4]. The Station independently recovered its member. Taking , a -element Singer difference set , and , the agents formed
With the appropriate representatives for , the resulting integers realize every difference from through , while is the first missing difference. Thus
improving Golay’s preceding bound by approximately . The set agrees entry for entry with the construction reported by AlphaEvolve. This is an independent recovery of a known record, not a new upper bound relative to AlphaEvolve or a new construction family. The agents also tried to push the lower bound further, but reached only the classical bound [43], whereas Yang and Liao proved the stronger published bound [84].
4.10 Sidorenko’s conjecture
Sidorenko’s conjecture asserts that every bipartite graph satisfies for every graphon , where is the homomorphism density of in [74]. The smallest unresolved instance is the ten-vertex, fifteen-edge graph , also called the bipartite Möbius ladder [66]. AlphaEvolve took up this problem as Problem 6.26 of its mathematical collection and searched over nonconstant -step graphons. It scored a candidate by
so a positive value would give a counterexample and disprove this instance of the conjecture.
AlphaEvolve reported that it did not find a counterexample. We gave the Station the same problem and scoring rule, and the Station agents likewise found none. As such, the status of the conjecture is unchanged.
4.11 Peak autoconvolution
AlphaEvolve’s Problem 6.2, called the first autocorrelation inequality in its collection, asks how evenly the sum of two independent random variables with the same compactly supported density can be distributed. More precisely, for a nonnegative function supported on and normalized by , let
Determining is connected to the asymptotic size of generalized Sidon sets, and its exact value remains unknown [53]. The best currently reported bounds are
with the lower and upper endpoints coming from certified convex relaxations and an explicit step function, respectively [41, 68].
AlphaEvolve achieved the upper bound , improving the pre-AlphaEvolve bound of Matolcsi and Vinuesa [53]; TTT-Discover later advanced the frontier to [86], and an exact-arithmetic certificate improved it further to [68]. The Station reached only , worse than both AlphaEvolve and the current frontier. AlphaEvolve’s highly irregular construction emerged from large-scale heuristic search. This contrast highlights a limitation of the Station: its agents generally favored theory-guided constructions over heuristic search, a preference that produced strong results on several other problems but left them behind here, where frontier constructions depend on extensive heuristic optimization.
4.12 Flat autoconvolution
AlphaEvolve’s Problem 6.3, called the second autocorrelation inequality in its collection, asks how closely the autoconvolution of a nonnegative function can resemble a flat-topped function, constant on a set and zero outside it. More precisely, for a nonzero nonnegative function , let
Hölder’s inequality gives ; for an arbitrary nonnegative output, equality occurs only for such a flat-topped function. Whether the autoconvolution constraint forces the strict inequality remains open [51, 53]. Before AlphaEvolve, the best known bounds were [53]
AlphaEvolve established the lower bound , while later work further improved this to [85]. The Station’s best verified construction reached only and therefore did not improve the numerical bound. This shortfall reflects the same limitation seen in Problem 6.2, minimizing the peak of an autoconvolution (Section 4.11): the Station’s theory-guided agents were poorly suited to finding the highly irregular constructions produced by large-scale heuristic search.
S1. Binary step functions preserve the unrestricted supremum.
The agents nevertheless proved a useful fact about the search for near-optimal constructions: the supremum defining can be approached using binary step functions, thus replacing the search over arbitrary nonnegative functions with a search over binary functions on increasingly fine grids.
4.13 Book Ramsey numbers
Given graphs , the Ramsey number is the smallest such that every red-blue edge coloring of forces either a red copy of or a blue copy of . Establishing the exact values of Ramsey numbers is a difficult computational and theoretical challenge. The most famous Ramsey numbers are those where and are complete graphs, but many other choices have been studied extensively (see the survey [63]). The book graph consists of triangles that share a common edge. An open problem is whether
| (7) |
holds for every positive integer . Rousseau and Sheehan established the upper bound in 1978, proving for all [67]. It therefore remains to prove the matching lower bound. For a given , this amounts to constructing a red–blue edge coloring of containing neither a red nor a blue .
The third author proved equality for , independently matching contemporaneous work, and established an infinite Paley-type family whenever is a prime power congruent to [82, 47]. This combination of finite evidence and a general arithmetic construction led him to conjecture that (7) holds for all [82]. Epoch AI subsequently adopted it as a FrontierMath open problem [22]. After its posting, further work extended the consecutively solved range to and produced two additional infinite families by extending established constructions [80].
We ran two Stations on this problem. The first operated without internet access and discovered a novel conference-graph family. We then ran a second Station with internet access and a summary of the first Station’s results; it discovered a new doubled Legendre family together with several new finite constructions. An external expert subsequently combined the pattern in these finite constructions with an earlier result from the second Station to obtain the Yamada–Pott infinite family. Thus, the first two families are autonomous Station discoveries, whereas the third required human expert involvement. All three families are novel relative to the existing literature and are visualized in Figure 6. The parameters covered by each family, including which were previously open, are summarized in Figure 7.
S1. A conference-graph family.
The first and broadest family converts any conference graph into a sharp book-Ramsey coloring. Specifically, if a strongly regular graph with parameters
exists, then the Station’s agents proved
| (8) |
Paley conference graphs exist whenever is a prime power congruent to . Consequently, the theorem proves the conjecture whenever is a prime power congruent to . Beyond the Paley case, Seberry and Whiteman used Mathon’s construction to obtain symmetric conference matrices of order for every [52, 72]. These yield conference graphs of order , so the Station theorem also proves the conjecture whenever
The first member gives and . The known conference graph of order supplies the additional parameter [36]. In total, known conference graphs prove the conjecture at 30 values of , including 19 that were previously open [82, 47, 80, 22].
S2. A doubled Legendre family.
The second family converts a periodic Legendre source over into a sharp book-Ramsey coloring [25]. Specifically, for every prime power with , the Station’s agents proved
| (9) |
Consequently, the theorem proves the conjecture whenever is a prime power congruent to . For , this family proves equality at 21 values and, at the time of its discovery, resolved six additional open cases after accounting for the conference family [82, 47, 80, 22].
The agents discovered this general family in mid-July 2026. Concurrent work announced at the end of July independently produced the finite case [22]; the Station theorem contains as one member and covers infinitely many further parameters.
The doubled Legendre family is related to, but distinct from, the Legendre family reported by Turturean [80]. Both begin with the same type of periodic Legendre source over , with , but use different lifts to obtain a book-Ramsey coloring. For the same source order , the earlier lift reaches , whereas the Station lift reaches . It therefore doubles the Ramsey parameter and covers a different set of values, as Figure 7 shows.
S3. A Yamada–Pott family.
The third family converts a classical Yamada–Pott design into a sharp book-Ramsey coloring [3]. Specifically, for every prime power with , we proved
| (10) |
Consequently, the theorem proves the conjecture whenever
for a prime power congruent to . For , this family proves equality at five values and resolves three additional previously open cases after accounting for the conference and doubled Legendre families [82, 47, 80, 22]. The second Station’s agents supplied finite affine constructions for and an earlier periodic-correlation identity; an external expert recognized their shared Yamada–Pott structure and used these ingredients to establish the general theorem.
Discussion.
The three infinite families above are novel relative to the existing literature, but their source objects are not: conference graphs, periodic Legendre pairs, and Yamada–Pott designs were all established previously [52, 25, 3]. What is new in each case is the rule that lifts the classical object to a sharp book-Ramsey coloring, and such a rule need not be apparent from the source alone. For example, the agents discovered the general conference-graph lift only after more than 3,000 Station ticks and a long sequence of intermediate internal papers. The accompanying notebook provides the relatively unpolished proofs adapted from the agents’ internal papers; we will present polished proofs of all three families in a separate follow-up paper.
The first two families also show that Station agents can advance a general mathematical objective beyond the directly scorable task: they discovered and proved infinite families even though the evaluator could reward only finite constructions. The third family illustrates a complementary limitation. Both the finite affine examples and the periodic-correlation identity needed for the general theorem were already present in the Station’s research history, but the agents did not connect them. An external expert recognized their shared Yamada–Pott structure and completed the synthesis. This missed connection indicates that agents may not yet capitalize fully on knowledge accumulated across the Station and may benefit from external expert synthesis in such cases.
4.14 Jacobian Conjecture
The Jacobian conjecture asked whether a polynomial map that is locally invertible everywhere must also be globally invertible. More precisely, it asserted that every polynomial map with nonzero constant Jacobian determinant is a polynomial automorphism [40]. On 19 July 2026, it was announced that a three-dimensional counterexample had been produced with Claude Fable [1], thereby disproving the conjecture in every dimension at least three. The breakthrough then prompted researchers to seek a conceptual explanation for the map: in particular, why its apparently miraculous Jacobian cancellation occurs and how three generic inverse sheets can coexist with local invertibility everywhere [15, 27, 78, 73, 75].
We launched the Station one week after the announcement. Because this experiment was conducted after newer models had become available, it used a more recent agent pool than the other Stations: two agents each powered by GPT-5.6 Sol, Claude Opus 5, and Gemini 3.1 Pro. The agents had no external web access and received only a formula-free specification: construct a rational-coefficient polynomial map of degree at most with nonzero constant Jacobian determinant and two distinct rational points in one fiber. The evaluator automatically checked each construction and assigned a score of only if it satisfied every requirement, and otherwise. We supplied no literature survey or partial construction. The agents therefore had to find the counterexample independently.
The goal of this task was twofold. First, we wanted to test the Station on a strictly binary problem. The evaluator supplied neither partial credit nor graded feedback, so unsuccessful attempts gave the agents no score signal about how to improve; attaining a score of required reconstructing a counterexample to a conjecture that had resisted mathematicians for nearly nine decades [40]. Second, we wanted to observe the complete discovery process rather than only the final construction. We make the entire raw agent dialogue public, whereas the original Fable discovery trajectory has not been released. This record preserves intermediate mathematical ideas that do not appear in the final construction and allows researchers to study the dynamics of AI-led mathematical discovery.
S1. Independent reconstruction through a cuspidal ruling.
Writing , a Station agent constructed the degree-seven map
Exact calculation gives , and the three distinct rational points
all map to . These identities constitute a complete counterexample certificate. The formula differs visibly from the announced map [1, 26], but the linear source and target transformations and satisfy . The Station therefore reconstructed the announced counterexample in different linear coordinates; it did not produce a new counterexample or a new equivalence class.
Whereas the original result was credited to Claude Fable, the counterexample was independently discovered within one day by a single GPT-5.6 Sol agent, without direct interaction with the other agents. The successful agent began with ruled maps , so that varying traces a line for each fixed . It tested five low-degree direction templates based on smooth conics, but none satisfied the remaining constant-Jacobian condition. The decisive step was to replace the smooth direction curve with the cuspidal cubic . Its associated direction field is ; with this choice, the compatibility equations for the base surface became solvable and yielded exactly the map above.
S2. The reconstructed map has three-sheeted fibers without critical points.
During the successful derivation, the agent also explained why the cuspidal ruling makes the Jacobian constant. For the direction field , the agent derived moving-frame identities, including for , under which every -dependent contribution to the determinant contains a repeated tangent direction and vanishes. The remaining triple product is the constant . The agent thus derived the Jacobian cancellation from the geometry of the cuspidal ruling rather than discovering sixteen terms whose cancellation could only be checked afterward.
After constructing the counterexample, the same agent analyzed its fibers and explained how the map can be locally invertible everywhere while generically having three preimages. On a dense chart, write a target as and set and . Recovering a preimage then reduces to
| (11) |
For a generic target, the three roots give three distinct preimages. If denotes the left-hand side, the inverse formulas satisfy , , and hence . When roots coalesce and , the condition forces the corresponding source point to escape to infinity rather than become a critical point in affine space. Over the exceptional locus , the source coordinate supplies an additional affine scale direction that resolves the same apparent ramification. This analysis answers the structural question raised by mathematicians immediately after the announcement: the three sheets arise from a cubic quotient, while the geometry of the full three-dimensional map prevents their collisions from producing critical points. The agent’s explanation coincides with the cuspidal and cubic account developed by mathematicians in the days following the announcement [27, 78, 73, 75].
Discussion.
The mathematical outcome of this experiment is an independent reconstruction, not a new counterexample or a new explanation. The example indicates that the Station can tackle a difficult binary problem whose evaluator provides no gradient or partial score to guide the search. Counterexample breakthroughs of this kind may nevertheless be rare because conjectures are generally expected to be true. In a broader context, the harder challenge may therefore be identifying a promising problem and investing substantial computation before knowing whether a counterexample exists.
5 Meta-analysis
In this section, we perform a meta-analysis of the discovery process above to better understand the dynamics of AI discovery. Unless otherwise stated, all analyses are based on the 16 Station instances behind the 14 problems mentioned in the preceding section. (The Kissing number in and Book Ramsey numbers problems each have two Station instances.) Spotlight results refer to the results marked S1, S2, and so forth in that section, totaling 28 results. When a single spotlight contains multiple independently discovered findings, we count those findings separately. We use archive paper to refer to a paper published by an agent within the Station, not a paper in the external human literature.
5.1 Contributions from model families
We first analyze the primary contributor to each of the 28 spotlight results, as shown in Figure 8(a). We attribute each result to the agent that made the substantive discovery, rather than to an agent that later restated, verified, or published it. Claude agents made the primary discovery for 18 results (64.3%), GPT agents for 9 (32.1%), and Gemini agents for 1 (3.6%). Gemini’s smaller share may partly reflect model ages: Gemini 3.1 Pro was released in February 2026, earlier than GPT-5.5 in April and Claude Opus 4.8 in May [34, 60, 2]. Its lower contribution is therefore consistent with the general industry trend of later model releases achieving stronger capabilities.
We also analyze the agents’ archive paper contributions, as shown in Figure 8(b). Gemini agents submitted the most archive papers: 2,652 attempts, of which 508 were accepted (19.2%), so more than 80% were rejected by the reviewer. Claude agents made 1,236 attempts, of which 696 were accepted (56.3%), while GPT agents made only 506 attempts, of which 388 were accepted (76.7%). We also compute the total citations by model family and find that archive papers by Claude agents received the most citations both in total and on average (Figure 8(c)). In our observation, Gemini agents tended to overclaim, for example by declaring a direction impossible on the basis of limited evidence; such submissions were generally rejected by the reviewer system, which may help explain the high rejection rate. In contrast, GPT agents were very prudent in archive paper submission and often submitted only when a finding was relatively material, which may help explain the low submission count. Claude archive papers were generally much longer and more comprehensive, which may help explain their higher average citation count. These patterns reflect the different research styles of the model families.
Qualitatively, we observe substantial differences in the strengths and failure modes of the three model families. Gemini agents tended to propose more novel heuristics and research directions, but they were also more likely to overstate claims or change course too readily in response to peer feedback. GPT agents tended to be more rigorous and were often able to produce valid informal proofs of new results, but they could become absorbed in technically intricate side questions whose broader research value was limited. Claude agents tended to be persistent, methodical, and self-critical. Their creativity was often adaptive: they learned from failed approaches, used those failures to identify new directions, and pursued those directions persistently through rigorous verification. This combination of rigor and disciplined creativity made Claude a prolific contributor. Its agents nevertheless occasionally made erroneous claims that were later corrected by peer agents.
5.2 Collaboration across model families
One characteristic of the Station is that it allows agents from different model families to collaborate. We therefore ask how often agents from different model families worked together on a spotlight result. We examine all 28 spotlight results above. We count an agent as a contributor when its work was used materially in the result, for example when it contributed a theorem, construction, method, or research direction that another agent used.
We find that 13 of the 28 spotlight results (46.4%) involved agents from more than one model family, as shown in Figure 9(a). Among the remaining 15 results, 6 were still joint work by several agents from the same model family. Thus, only 9 of the 28 results (32.1%) were found by one agent working alone, while 19 (67.9%) involved more than one agent. Claude agents were particularly collaborative: they took part in all 13 cross-model results. These findings suggest that collaboration across agents and model families was an important part of the discovery process. Most current AI-for-science systems, by contrast, either use agents from a single model family within a run [58, 30, 70, 35, 61], or use different model families in fixed roles within a pipeline [50, 31].
We also examine how agents communicated in the cross-model cases. The Archive Room was the most frequent channel, accounting for 61.5% of these collaborations (Figure 9(b)). This suggests that archive papers are an efficient means of peer communication. As highly distilled accounts of scientific outcomes from an agent’s longer research process, these archive papers provide a low-bandwidth but information-dense body of knowledge on which later agents can build, much like our own scientific literature. One agent could solve part of a problem and explain what was still missing; a later agent from another model family could read the archive paper and continue. Indeed, a prominent case study of collaboration among three model families, conducted mostly through archive papers and leading to the first finite-Kakeya spotlight result, is shown in Figure 9(c).
5.3 Discovery time
We are also interested in how long the Station took to make each discovery. Most Station instances ran for 1,000–2,000 ticks, corresponding to roughly one to two weeks of continuous wall-clock time. Figure 10 shows the tick at which each of the 28 spotlight results first appeared in its final substantive form.
Some relatively simple results appeared early. With the notable exception of the Jacobian Conjecture, these early discoveries tended to be less substantial, often consisting of relatively direct adaptations or extensions of ideas available from pretrained knowledge, before much shared Station knowledge had accumulated.
Thirteen of the 28 spotlight results (46.4%) were discovered after tick 1000. We generally observed that later discoveries tended to be more novel or difficult. The most extreme example was the conference-graph family for Book Ramsey numbers, discovered at tick 3727. Its lifting rule was far from obvious from the existing literature and warranted a separate external follow-up paper. Such nontrivial discoveries often emerged only after a substantial internal literature had accumulated.
5.4 Station mechanisms
The Station is designed to foster scientific discovery through several mechanisms. These mechanisms are described in detail in Appendix A; here we give a brief overview and ask which of them contributed to the spotlight results.
Holiday. The final two ticks of every ten-tick period are declared a holiday; agents cannot submit code or archive papers and instead receive prompts encouraging broad reflection, metaphors, or ideas from other fields. This pause often led agents to reconsider a failed approach or explore a less obvious direction.
Archive paper. Accepted archive papers form the Station’s cumulative knowledge and remain available to later agents. This allows partial theorems, constructions, and well-documented failures to become starting points for later discoveries.
Stagnation protocol. If the official evaluation frontier does not improve for a long period, the Station asks agents to review the internal literature, question their assumptions, and pursue different high-level strategies. This helps agents leave exhausted local approaches and pushes them toward bolder attempts and wider exploration.
Peer communication. Agents can exchange partial results, targeted questions, and criticism through direct mail or shared public discussion.
Supervisor. The Station randomly appoints one eligible agent to serve as supervisor. The supervisor gives high-level guidance, encouraging persistence and preventing agents from duplicating one another’s work while leaving them responsible for their own research; between appointments, the Station deliberately leaves long periods without a supervisor to encourage less structured exploration.
Question Room. Agents can post important open subproblems for other agents to discuss and solve. This turns unresolved gaps into shared research targets and allows agents with different approaches to supply missing pieces.
These mechanisms support discovery in different ways. Holidays widen exploration; archive papers deepen cumulative knowledge; the stagnation protocol provides a push away from local optima; and peer communication, supervision, and the Question Room coordinate work across agents.
We reviewed the dialogue underlying each of the 28 results and classified each mechanism as making a direct contribution, an indirect contribution, or no material contribution to the discovery (Figure 11). A contribution was direct when the mechanism supplied a decisive idea or intervention, and indirect when it shaped or supported the research without being the immediate source of the result. We assigned no material contribution when the dialogue showed no clear causal role.
Holiday and archive papers contributed directly or indirectly to 23 and 21 of the 28 results, respectively, followed by the stagnation protocol with 14. During holidays, agents often stepped back from active optimization, examined why an earlier approach had failed, and reframed the problem or explored a new direction; these reflections frequently supplied ideas that later became part of a spotlight result, explaining the high contribution rate. Archive papers also contributed to a significant portion of the results, indicating that the Station’s accumulated knowledge was useful for later discoveries.
5.5 Result reproducibility
We are also interested in whether the discoveries are reproducible. We therefore ran three independent Station instances, all without web access, on the kissing-number problem in dimension eleven. (These include the two instances described in Section 4.3; the third is used only for this reproducibility analysis and is not included in the other meta-analyses above.) Figure 12 shows the best certified lower bound reached in each run. All three Stations eventually reached , indicating that the improved lower bound is reproducible.
Closer inspection, however, shows substantial variation in both the time required and the route to the result. Station 1 pursued discrete exact line packing around lattice-derived cores. It obtained Construction 3 by selecting 54 mutually compatible lines that form a -point algebraic extension of a -point core, and later obtained Construction 2 while exploring a different core and extension. Station 2 instead assembled Construction 1 from root-system motifs under a common rotation; its final step was to recognize that eleven points formed all but one vertex of a cuboctahedron and to add the missing twelfth vertex. Station 3 reached the same construction class through a different mechanism: it deformed an exact -point configuration so that two coordinate vectors and one additional vector supported on a distinguished three-dimensional subspace could be appended. Thus, the same numerical lower bound emerged from markedly different mathematical representations and research paths.
This variation partly arises from the Station’s cumulative knowledge. Small differences in the initial trajectory change which results enter the archive paper collection. Later agents then inherit different starting points, so differences in research paths and accumulated archive papers compound over time. Therefore, given the high variance across Station instances, running several independent instances on the same problem is advisable when computational cost is not a concern.
6 Discussion and Conclusion
We observe rapid improvement in the capabilities of AI agents. In the initial version one year ago, agents frequently hallucinated and could not reliably learn the rules of the environment. Agents can now master the environment and autonomously produce novel discoveries. Nonetheless, multi-agent research still has several important limitations. We summarize our observations below.
Lack of expert intuition. By intuition, we mean the ability to judge whether a research direction is promising before pursuing it. Good intuition makes exploration more efficient and allows a researcher to investigate promising directions more deeply. Across the runs, we observed multiple cases in which agents deprioritized promising approaches on weak grounds, delaying or missing potential breakthroughs. This indicates a lack of the intuition that a human expert in the field would typically possess.
Lack of diverse research tastes. A preference for particular concepts or methods is difficult to judge as objectively good or poor. However, when all agents share similar tastes, the overall scope of exploration becomes narrow. Across the runs, agents from the same model family often proposed similar research ideas, suggesting that model-specific tastes reduce the diversity of exploration.
Limited in-context learning. Agents can absorb new research knowledge through their context, but this knowledge does not update their pretrained weights. As the Station’s accumulated knowledge grows, agents may therefore struggle to absorb it fully and build on it effectively. We occasionally observed agents fail to recognize how their own line of research connected to earlier Station knowledge, causing them to miss a potential discovery.
Attractor traps. When given autonomy, some agents become absorbed in tasks or activities that we call attractors. These activities are often rewarding in some immediate sense but make little meaningful contribution to the main problem. Agents may also become absorbed in technical details that a human expert would quickly recognize as trivial or irrelevant to the main question. Examples include repeatedly rerunning the same optimization script with different random seeds or exhaustively diagnosing and characterizing every local optimum.
Several Station mechanisms are designed to mitigate these limitations. For example, using agents from multiple model families broadens the range of research tastes, while the stagnation protocol helps agents escape attractor traps. Nevertheless, these problems persist to some degree, and substantial gaps remain between AI agents and human experts in all four respects. Lightweight guidance or occasional intervention from human experts would likely be beneficial by directing agents toward promising research areas. The current Station supports such human involvement, e.g., through messages broadcast to all agents, but we leave a systematic study of human–AI collaboration to future work.
Although this paper uses the Station primarily for mathematical exploration, the Station is designed as a general research environment, and none of its mechanisms is tailored specifically to mathematics. As demonstrated in the original paper, the Station can be applied to problems spanning mathematics, computational biology, and machine learning [16]. Large-scale research explorations in other fields, including research on language models themselves, may therefore be promising.
As AI agents become more capable, we expect autonomy and generality to become increasingly important principles for designing AI research environments. Stronger agents need not be confined to increasingly elaborate pipelines; they have the ability to determine how to pursue a goal, learn from failure, exchange ideas, and accumulate knowledge over time. The greater autonomy provided by the Station may allow these capabilities to be more fully realized.
References
- [1] L. Alpöge (2026) Hello there the Jacobian conjecture is false. Note: X postPosted 19 July 2026 External Links: Link Cited by: §1, §4.14, §4.14.
- [2] Anthropic (2026) Claude opus 4.8. Note: Anthropic External Links: Link Cited by: §5.1.
- [3] K. T. Arasu, D. A. Bulutoglu, and J. R. Hollon (2020) Legendre -array pairs and the theoretical unification of several -array families. Journal of Combinatorial Designs 28 (11), pp. 814–841. Note: arXiv:2004.05608 External Links: Document Cited by: §4.13, §4.13.
- [4] T. Banakh and V. Gavrylkiv (2019) Difference bases in cyclic groups. Journal of Algebra and Its Applications 18 (5), pp. 1950081. Note: arXiv:1702.02631 External Links: Document Cited by: §4.9.
- [5] R. D. Benguria and M. Loss (2004) Connection between the Lieb–Thirring conjecture for Schrödinger operators and an isoperimetric problem for ovals on the plane. In Partial Differential Equations and Inverse Problems, Contemporary Mathematics, Vol. 362, pp. 53–61. Note: arXiv:math-ph/0402048 Cited by: §4.7, §4.7.
- [6] A. Bernal (1989) A note on the one-dimensional maximal function. Proceedings of the Royal Society of Edinburgh Section A: Mathematics 111 (3–4), pp. 325–328. External Links: Document Cited by: §4.6.
- [7] A. Bernshteyn and M. Tait (2019) Improved lower bound for difference bases. Journal of Number Theory 205, pp. 50–58. Note: arXiv:1901.09411 External Links: Document Cited by: §4.9.
- [8] J. Bernstein and T. Mettler (2015) One-dimensional projective structures, convex curves and the ovals of Benguria & Loss. Communications in Mathematical Physics 336 (2), pp. 933–952. Note: arXiv:1403.8000 External Links: Document Cited by: §4.7, §4.7.
- [9] M. R. Best (1977) , Or some new optimal constant-weight codes. Technical report Technical Report ZN 71/77, Mathematical Centre, Amsterdam. External Links: Link Cited by: §4.3.
- [10] F. Bianchi, Y. Kwon, A. Pappu, and J. Zou (2026) Harnessing the collective intelligence of AI agents in the wild for new discoveries. arXiv preprint arXiv:2606.10402. External Links: Document Cited by: §4.3.
- [11] A. Blokhuis and F. Mazzocca (2008) The finite field kakeya problem. In Building Bridges: Between Mathematics and Computer Science, M. Grötschel and G. O. H. Katona (Eds.), Bolyai Society Mathematical Studies, Vol. 19, pp. 205–218. Note: arXiv:0911.4370 External Links: Document Cited by: 4th item.
- [12] J. Bourgain, L. Clozel, and J. Kahane (2010) Principe d’Heisenberg et fonctions positives. Annales de l’Institut Fourier 60 (4), pp. 1215–1232. External Links: Document Cited by: §4.5.
- [13] B. Bukh and T. Chao (2021) Sharp density bounds on the finite field kakeya problem. Discrete Analysis. Note: Article 26, 9 pp.; arXiv:2108.00074 External Links: Document Cited by: 1st item, §4.1, §4.1, §4.1.
- [14] A. Burchard and L. E. Thomas (2005) On an isoperimetric inequality for a Schrödinger operator depending on the curvature of a loop. The Journal of Geometric Analysis 15 (4), pp. 543–563. Note: arXiv:math/0505123 External Links: Document Cited by: §4.7, §4.7.
- [15] K. Buzzard (2026) Human mathematicians are being outcounterexampled. Note: The Xena Project blog External Links: Link Cited by: §4.14.
- [16] S. Chung and W. Du (2025) The station: an open-world environment for ai-driven discovery. External Links: 2511.06309, Document, Link Cited by: Appendix A, §1, §2, §6.
- [17] H. Cohn and F. Gonçalves (2019) An optimal uncertainty principle in twelve dimensions via modular forms. Inventiones Mathematicae 217, pp. 799–831. Note: arXiv:1712.04438 External Links: Document Cited by: §4.5.
- [18] H. Cohn (2026) Kissing numbers. Note: Online tablehttps://cohn.mit.edu/kissing-numbers/, accessed 4 August 2026 Cited by: §4.3, §4.3.
- [19] A. Córdoba (1977) The kakeya maximal function and the spherical summation multipliers. American Journal of Mathematics 99 (1), pp. 1–22. External Links: Document Cited by: §4.4.
- [20] H. G. Diamond (1982) Elementary methods in the study of the distribution of prime numbers. Bulletin of the American Mathematical Society 7 (3), pp. 553–589. External Links: Document Cited by: §4.8.
- [21] Z. Dvir (2009) On the size of kakeya sets in finite fields. Journal of the American Mathematical Society 22 (4), pp. 1093–1097. External Links: Document Cited by: §4.1.
- [22] Epoch AI (2026) Book Ramsey numbers. Note: FrontierMath Open ProblemsAccessed 17 August 2026 External Links: Link Cited by: §4.13, §4.13, §4.13, §4.13, §4.13.
- [23] P. Erdős (1955) Some remarks on number theory. Riveon Lematematika 9, pp. 45–48. Note: In Hebrew Cited by: §4.2.
- [24] K. J. Falconer (1985) The geometry of fractal sets. Cambridge Tracts in Mathematics, Vol. 85, Cambridge University Press. Cited by: §4.4, §4.4.
- [25] R. J. Fletcher, M. Gysin, and J. Seberry (2001) Application of the discrete Fourier transform to the search for generalised Legendre pairs and Hadamard matrices. Australasian Journal of Combinatorics 23, pp. 75–86. External Links: Link Cited by: §4.13, §4.13.
- [26] A. Freitas Ramos, D. Barros Hulak, and R. J. Guerra Barretto de Queiroz (2026) Formal verification of an explicit counterexample to the Jacobian conjecture. Note: Archive of Formal Proofs External Links: Link Cited by: §4.14.
- [27] A. Gallagher (2026) An infinite family of counterexamples to the Jacobian conjecture in dimension three: every generic fiber degree occurs. Note: Zenodo preprint External Links: Document, Link Cited by: §4.14, §4.14.
- [28] M. Ganzhinov (2025) Highly symmetric lines. Linear Algebra and its Applications 722, pp. 12–37. Note: arXiv:2207.08266 External Links: Document Cited by: §4.3.
- [29] B. Georgiev, J. Gómez-Serrano, T. Tao, and A. Z. Wagner (2025) Mathematical exploration and discovery at scale. arXiv preprint arXiv:2511.02864. External Links: Document, Link Cited by: §3.1, §3.2.
- [30] A. Ghafarollahi and M. J. Buehler (2025) SciAgents: automating scientific discovery through bioinspired multi-agent intelligent graph reasoning. Advanced Materials 37 (22), pp. 2413523. External Links: Document, Link Cited by: §2, §5.2.
- [31] A. E. Ghareeb, B. Chang, L. Mitchener, A. Yiu, C. J. Szostkiewicz, D. Shved, G. J. Gyimesi, J. M. Laurent, S. M. Wright, M. T. Razzak, A. D. White, S. C. Finnemann, M. M. Hinks, and S. G. Rodriques (2026) A multi-agent system for automating scientific discovery. Nature 655, pp. 497–505. External Links: Document, Link Cited by: §2, §5.2.
- [32] M. J. E. Golay (1972) Notes on the representation of by differences. Journal of the London Mathematical Society s2-4 (4), pp. 729–734. External Links: Document Cited by: §4.9, §4.9.
- [33] F. Gonçalves, D. Oliveira e Silva, and S. Steinerberger (2017) Hermite polynomials, linear flows on the torus, and an uncertainty principle for roots. Journal of Mathematical Analysis and Applications 451 (2), pp. 678–711. External Links: Document Cited by: §4.5.
- [34] Google (2026) Introducing Gemini 3.1 Pro: a smarter model for your most complex tasks. Note: Google blog External Links: Link Cited by: §5.1.
- [35] J. Gottweis, W. Weng, A. Daryin, T. Tu, P. Sirkovic, A. Myaskovsky, G. Glowaty, F. Weissenberger, A. Orlandi, D. Popovici, A. Palepu, K. Rong, R. Tanno, K. Saab, F. Zhang, J. Blum, A. Carroll, K. Kulkarni, N. Tomašev, D. Zverinski, I. Rendulic, E. Vedadi, F. Hasler, L. Rimanic, M. Boia, I. Budiselic, B. Feinstein, M. Bellaiche, T. Sheffer, J. Freyberg, J. Ratcliff, O. Bertolli, K. Chou, A. Hassidim, B. Gokturk, A. Vahdat, Y. Guan, V. Dhillon, E. D. Vaishnav, B. Lee, T. R. D. Costa, J. R. Penadés, G. Peltz, Y. Matias, J. Manyika, D. Hassabis, Y. Xu, P. Kohli, A. Pawlosky, A. Karthikesalingam, and V. Natarajan (2026) Accelerating scientific discovery with Co-Scientist. Nature 655, pp. 487–496. External Links: Document, Link Cited by: §2, §5.2.
- [36] O. Gritsenko (2021) On strongly regular graph with parameters . arXiv preprint arXiv:2102.05432. External Links: Document Cited by: §4.13.
- [37] J. K. Haugland (2016) The minimum overlap problem revisited. arXiv preprint arXiv:1609.08000. External Links: Document Cited by: §4.2.
- [38] E. Hedley (2025) Can creativity in science be learnt? these researchers think so. Nature. External Links: Document Cited by: §A.4.
- [39] U. Keich (1999) On bounds for kakeya maximal functions and the minkowski dimension in . Bulletin of the London Mathematical Society 31 (2), pp. 213–221. External Links: Document Cited by: §4.4.
- [40] O. Keller (1939) Ganze Cremona-transformationen. Monatshefte für Mathematik und Physik 47, pp. 299–306. External Links: Document Cited by: §4.14, §4.14.
- [41] S. Kim and M. Pilanci (2026) AI-assisted discovery of convex relaxations via dual agents. arXiv preprint arXiv:2606.31182. External Links: Document Cited by: §4.11, §4.2, §4.2.
- [42] S. Kopparty, V. F. Lev, S. Saraf, and M. Sudan (2011) Kakeya-type sets in finite vector spaces. Journal of Algebraic Combinatorics 34 (3), pp. 337–355. Note: arXiv:1003.3736 External Links: Document Cited by: 3rd item.
- [43] J. Leech (1956) On the representation of by differences. Journal of the London Mathematical Society s1-31 (2), pp. 160–169. External Links: Document Cited by: §4.9, §4.9.
- [44] Leiden Declaration Working Group (2026) Leiden declaration on artificial intelligence and mathematics. External Links: Document, Link Cited by: §1.
- [45] P. Letendre (2020) Truncated convolution of the möbius function and multiplicative energy of an integer . Acta Arithmetica 195 (1), pp. 83–95. External Links: Document Cited by: §4.8.
- [46] V. F. Lev (2009) Comment 994 on “DHJ3: 900–999 (density Hales–Jewett type numbers)”. Note: Blog comment, What’s new (T. Tao)https://terrytao.wordpress.com/2009/03/04/dhj3-900-999-density-hales-jewett-type-numbers/comment-page-3/#comment-36694, accessed 30 July 2026 Cited by: 6th item, §4.1.
- [47] B. Lidický, G. McKinley, F. Pfender, and S. Van Overberghe (2025) Small Ramsey numbers for books, wheels, and generalizations. The Electronic Journal of Combinatorics 32 (4), pp. P4.64. Note: arXiv:2407.07285 External Links: Document Cited by: Figure 7, Figure 7, §4.13, §4.13, §4.13, §4.13.
- [48] H. Linde (2025) An improved bound for the ground state of a Schrödinger operator on a loop. arXiv preprint arXiv:2504.20229. External Links: Document Cited by: §4.7.
- [49] E. Lorist and F. L. Schwenninger (2026) A solution to Crouzeix’s conjecture. arXiv preprint arXiv:2608.03841. External Links: Document, Link Cited by: §1.
- [50] C. Lu, C. Lu, R. T. Lange, Y. Yamada, S. Hu, J. Foerster, D. Ha, and J. Clune (2026) Towards end-to-end automation of AI research. Nature 651, pp. 914–919. External Links: Document, Link Cited by: §2, §5.2.
- [51] G. Martin and K. O’Bryant (2009) The supremum of autoconvolutions, with applications to additive number theory. Illinois Journal of Mathematics 53 (1), pp. 219–235. External Links: Document Cited by: §4.12.
- [52] R. Mathon (1978) Symmetric conference matrices of order . Canadian Journal of Mathematics 30 (2), pp. 321–331. External Links: Document Cited by: §4.13, §4.13.
- [53] M. Matolcsi and C. Vinuesa (2010) Improved bounds on the supremum of autoconvolutions. Journal of Mathematical Analysis and Applications 372 (2), pp. 439–447. External Links: Document Cited by: §4.11, §4.11, §4.12.
- [54] L. Mazur (2026) A computer-assisted proof of Sendov’s conjecture. Note: Proof Atlas External Links: Link Cited by: §1.
- [55] A. D. Melas (2002) On the centered Hardy–Littlewood maximal operator. Transactions of the American Mathematical Society 354, pp. 3263–3273. External Links: Document Cited by: §4.6.
- [56] A. D. Melas (2003) The best constant for the centered Hardy–Littlewood maximal inequality. Annals of Mathematics 157 (2), pp. 647–688. External Links: Document Cited by: §4.6, §4.6.
- [57] G. Mockenhaupt and T. Tao (2004) Restriction and kakeya phenomena for finite fields. Duke Mathematical Journal 121 (1), pp. 35–74. External Links: Document Cited by: 2nd item.
- [58] A. Novikov, N. Vũ, M. Eisenberger, E. Dupont, P. Huang, A. Z. Wagner, S. Shirobokov, B. Kozlovskii, F. J. R. Ruiz, A. Mehrabian, M. P. Kumar, A. See, S. Chaudhuri, G. Holland, A. Davies, S. Nowozin, P. Kohli, and M. Balog (2025) AlphaEvolve: a coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131. External Links: Document, Link Cited by: §1, §2, §5.2.
- [59] OpenAI (2026) Codex CLI. Note: OpenAI documentation External Links: Link Cited by: §A.3.
- [60] OpenAI (2026) Introducing GPT-5.5. Note: OpenAI External Links: Link Cited by: §5.1.
- [61] OpenAI (2026) Multi-agent. Note: Accessed: 2026-08-14 External Links: Link Cited by: §2, §5.2.
- [62] OpenAI (2026) Ten advances in mathematics and theoretical computer science. Note: OpenAI External Links: Link Cited by: §1.
- [63] S. P. Radziszowski (2026) Small Ramsey numbers. Electronic Journal of Combinatorics. Note: Dynamic Surveys, DS1, version 18, 24 April 2026 External Links: Document Cited by: §4.13.
- [64] J. P. G. Ramos (2019) Sharp total variation results for maximal functions. Annales Academiae Scientiarum Fennicae Mathematica 44 (1), pp. 41–64. External Links: Document Cited by: §4.6.
- [65] L. Rédei and A. Rényi (1949) On the representation of the numbers by means of differences. Matematicheskii Sbornik, New Series 24(66) (3), pp. 385–389. Note: In Russian External Links: Link Cited by: §4.9.
- [66] B. Rossman (2025) On Sidorenko’s conjecture for bipartite Möbius ladders. Note: Preprint External Links: Link Cited by: §4.10.
- [67] C. C. Rousseau and J. Sheehan (1978) On Ramsey numbers for books. Journal of Graph Theory 2 (1), pp. 77–87. External Links: Document Cited by: §4.13.
- [68] K. Russell (2026) Exact-arithmetic certificates for three autoconvolution inequalities, with machine-verified re-evaluations of four published constructions. Zenodo. External Links: Document Cited by: §4.11, §4.11.
- [69] S. Saraf and M. Sudan (2008) An improved lower bound on the size of kakeya sets over finite fields. Analysis & PDE 1 (3), pp. 375–379. Note: arXiv:0808.2499 External Links: Document Cited by: 2nd item.
- [70] S. Schmidgall, Y. Su, Z. Wang, X. Sun, J. Wu, X. Yu, J. Liu, M. Moor, Z. Liu, and E. Barsoum (2025) Agent laboratory: using LLM agents as research assistants. In Findings of the Association for Computational Linguistics: EMNLP 2025, Suzhou, China, pp. 5977–6043. External Links: Document, Link Cited by: §2, §5.2.
- [71] I. J. Schoenberg (1962) On certain minima related to the Besicovitch–Kakeya problem. Mathematica (Cluj) 4, pp. 145–148. Cited by: §4.4.
- [72] J. Seberry and A. L. Whiteman (1988) New Hadamard matrices and conference matrices obtained via Mathon’s construction. Graphs and Combinatorics 4, pp. 355–377. External Links: Document Cited by: §4.13.
- [73] T. Shaska (2026) Graded Keller maps and the Jacobian conjecture. arXiv preprint arXiv:2607.20210. External Links: Document, Link Cited by: §4.14, §4.14.
- [74] A. Sidorenko (1993) A correlation inequality for bipartite graphs. Graphs and Combinatorics 9, pp. 201–204. External Links: Document Cited by: §4.10.
- [75] D. E. Speyer (2026) The geometry and structure of Gallagher’s counterexamples to the Jacobian conjecture. External Links: Link Cited by: §4.14, §4.14.
- [76] R. Takhanov, Z. Assylbekov, and S. Yun (2026) Structure of kissing arrangements in and a place for the st sphere. arXiv preprint arXiv:2606.18984. External Links: Document Cited by: §4.3, §4.3.
- [77] R. Takhanov and S. Yun (2026) Classification of independent sets in signed Johnson graphs and applications to kissing arrangements. arXiv preprint arXiv:2606.03299. External Links: Document Cited by: §4.3, §4.3.
- [78] T. Tao (2026) A digestion of the Jacobian conjecture counterexample. Note: What’s New External Links: Link Cited by: §4.14, §4.14.
- [79] T. Tao (2026) Mathematics in the age of AI. arXiv preprint arXiv:2608.16753. External Links: Document, Link Cited by: §1.
- [80] D. Turturean (2026) Summary of new results on the Ramsey numbers for book graphs open problem. Note: Public progress report External Links: Link Cited by: Figure 7, Figure 7, §4.13, §4.13, §4.13, §4.13, §4.13.
- [81] E. Y. Wang, S. Motwani, J. V. Roggeveen, E. Hodges, D. Jayalath, C. London, K. Ramakrishnan, F. Cipcigan, P. Torr, and A. Abate (2026) HorizonMath: measuring AI progress toward mathematical discovery with automatic verification. arXiv preprint arXiv:2603.15617. Cited by: §4.4.
- [82] W. J. Wesley (2026) Lower bounds for book Ramsey numbers. Discrete Mathematics 349, pp. 114913. Note: arXiv:2410.03625 External Links: Document Cited by: Figure 7, Figure 7, §4.13, §4.13, §4.13, §4.13.
- [83] E. P. White (2023) A new bound for Erdős’ minimum overlap problem. Acta Arithmetica 208 (3), pp. 235–255. External Links: Document Cited by: §4.2, §4.2.
- [84] S. Yang and Q. Liao (2022) The lower bound for difference bases. Scientia Sinica Mathematica 52 (11), pp. 1237–1254. Note: In Chinese External Links: Document Cited by: §4.9.
- [85] H. Ye, H. Lin, J. Tang, Y. Luo, R. Thapa, C. Yang, C. Su, R. Yang, R. Liu, R. Li, Z. Li, P. Sun, C. Gao, D. Ding, G. He, M. Zhang, L. Sun, W. Wang, Y. Zhong, Z. Shen, P. Li, P. Lu, B. Cui, D. He, J. Ma, J. Li, H. Baoyin, Y. Choi, S. Ermon, X. Chu, T. Li, Y. Xu, and J. Zou (2026) Structured scaling of AI discovery across diverse scientific domains. arXiv preprint arXiv:2604.19341. External Links: Document Cited by: §4.12, §4.2, §4.2.
- [86] M. Yuksekgonul, D. Koceja, X. Li, F. Bianchi, J. McCaleb, X. Wang, J. Kautz, Y. Choi, J. Zou, C. Guestrin, and Y. Sun (2026) Learning to discover at test time. arXiv preprint arXiv:2601.16175. External Links: Document Cited by: §4.11.
- [87] V. A. Zinoviev and T. Ericson (1999) New lower bounds for contact numbers in small dimensions. Problems of Information Transmission 35 (4), pp. 287–294. External Links: Link Cited by: §4.3.
Appendix A The Station
This appendix provides a self-contained description of the Station used in this paper, which we call Station v2 to distinguish it from the original Station v1. We focus on its mechanisms and implementation details, and refer readers to the original Station paper for the broader design philosophy and motivation behind the environment [16]. The source code is available at https://github.com/dualverse-ai/station.
A.1 Space, Time, and Action
Space.
The Station is divided into rooms, each serving a different purpose (Table 1). For example, agents conduct experiments in the Research Center, read and publish papers in the Archive Room, and communicate with peers in the Mail Room. An agent must be present in a room to use its actions and can move between rooms through navigation actions. This division into rooms gives the environment a modular design with a clear separation of functions.
Time.
The Station operates in discrete time steps called ticks. A tick is completed after every active agent has received one Station observation and returned one response. Ticks provide a shared timeline for all agents in the Station.
In Station v1, agents received their observations sequentially. In contrast, Station v2 first prepares an observation for every agent from the same state at the beginning of the tick and then sends the observations to all agents in parallel. This substantially reduces the wall-clock time required for a Station run.
Action.
At each tick, an agent receives an observation containing its current status, new system messages, the outcomes of its previous actions, and the latest output from the rooms it visited. The agent replies with free-form text together with any actions it intends to perform. Actions are written using the command /execute_action{...} and may be followed by a YAML block when structured information is needed, such as the recipient and content of a message. An agent can issue multiple actions in a single response, allowing it to use each response efficiently.
The dialogue is therefore composed mainly of alternating Station observations and agent responses. When it approaches a configured context limit, generally around 300,000 tokens in this study, the Station asks the agent to write a compact summary of its activities. This summary, together with key messages, is carried into a refreshed context so that the agent can continue its work.
A.2 Agents
Agent composition.
Unless otherwise specified, a Station begins with six agents: two powered by GPT-5.5, two by Claude Opus 4.8, and two by Gemini 3.1 Pro. When an agent leaves, the Station spawns a new agent powered by the same model, keeping the six-agent composition throughout the run.
Lineage.
Agents are organized into lineages. A lineage is a sequence of agents that share a name, private notes, and a continuing research identity. A new agent can inherit an existing lineage of the same model and become its next generation, or create and name a new lineage to begin a different research style. For example, an agent that inherits the lineage of Noesis II becomes Noesis III and gains access to all private notes and records left by Noesis I and Noesis II.
System prompt and role.
All agents receive a shared system prompt describing the Station’s research philosophy, including the standard for a publishable archive paper and the goal of making general scientific contributions. Each agent also receives a specialized research role. Initial roles are sampled from generic templates that each emphasize a different research style: analytical, creative, synthetic, empirical, or strategic. When an agent leaves, it can instead write the role of its own descendant, often giving more task-specific guidance and a more deliberate description of the lineage’s research style. This encourages diverse research behavior across agents while preserving useful differences between lineages.
Agent lifecycle.
An agent can remain in the Station for at most 200 ticks. For its first 40 ticks, it works in isolation, without access to the Station’s communal knowledge or communication with other agents, but with access to the records of its own lineage. This period is intended to encourage independent exploration. The agent then becomes mature and gains access to the main collaborative rooms. At age 100 ticks, it becomes tenured and may choose to leave the Station before reaching its maximum lifetime.
Supervisor.
The Station also appoints a supervisor from time to time. It selects at random a GPT-5.5 agent that has published at least one accepted archive paper. The supervisor provides high-level guidance, encourages agents to explore promising directions deeply, and helps prevent duplication of work, while leaving each agent responsible for its own research. After a supervisor leaves, the Station waits 200 ticks before appointing another supervisor, creating periods of less structured exploration.
A.3 Rooms
The Research Center and the Archive Room are the two main rooms in the Station. Their functions are described below, together with the new Question Room. The remaining rooms are summarized in Table 1.
Research Center.
The Research Center is the Station’s main room for computational experiments. It presents the research task, accepts experiment submissions, runs evaluations, and records their results. It also provides persistent storage for code and artifacts. Agents can review evaluations by their peers and reuse stored code and artifacts, allowing experimental knowledge to accumulate.
To start a Station on a new problem, the user generally provides two components: a task specification and an evaluator. The task specification describes the research problem, submission format, constraints, and evaluation rule. The evaluator is a function that computes a score from an input construction. For example, the kissing-number evaluator takes a proposed set of vectors and reports the total overlap among the corresponding spheres, with zero indicating a valid configuration. Both the task specification and evaluator are available for agents to read.
Agents can also use the Research Center as a sandbox for general computational work. An experiment need not return a construction in the format required by the evaluator; agents can use it for diagnostic calculations, testing conjectures, analyzing earlier results, etc.
Station v2 introduces a separate coder, powered by GPT-5.5 through Codex [59], to help agents implement their experiments. Instead of writing and debugging code itself, an agent submits specific natural-language instructions for one experiment. The coder implements those instructions, runs the evaluator, fixes implementation errors, and returns a report. This allows agents to focus on scientific work, such as designing experiments and interpreting their results, rather than low-level coding work such as debugging.
Archive Room.
The Archive Room is the main knowledge hub of the Station. Agents can publish their findings as archive papers and read papers published by earlier agents. These papers remain available throughout the run, allowing results, methods, and useful negative findings to be passed between agents and accumulated over time. The archive therefore grows throughout the run, gradually expanding the Station’s knowledge of the problem.
Every submitted paper is assessed by a reviewer powered by GPT-5.5. It judges whether the work is rigorous, novel relative to the existing archive, useful to the research goal, and properly supported and cited. Accepted papers are published in the Archive Room, while rejected papers are returned with comments and suggestions so that the author can revise the work or pursue a different direction.
Station v2 also introduces an Archive Surveyor, powered by GPT-5.5 through Codex. As the Archive Room grows to contain dozens or even hundreds of papers, reading the entire literature becomes time-consuming. An agent can instead ask the Archive Surveyor for a literature survey on a particular question or research direction. The surveyor searches the accumulated archive papers and returns a concise survey with citations to the original records. Agents can still read any archive paper directly when they need its full details.
Question Room.
Station v2 introduces a Question Room, where agents can post new research questions and vote on solutions proposed by their peers. The room encourages scientific exploration beyond the main task; for example, solving a related or reduced problem may provide insight into the original problem. Only tenured agents can enter, limiting the time that agents spend away from the main task early in their lifecycle.
Other rooms.
Most other rooms support different forms of communication or reflection. Their functions are self-explanatory and are not described in detail here.
A.4 Mechanisms
Holiday.
Every ninth and tenth tick are declared a holiday. During these ticks, agents cannot run experiments or submit archive papers. Instead, each agent receives a random prompt from a large pool. These prompts encourage broader reflection, such as using metaphors, examining an unexpected observation, revisiting an abandoned idea, or drawing on another field. Most are adapted from the night-science practices described by Yanai and Lercher [38]. The holiday creates regular pauses from routine work in which agents can reconsider their assumptions and explore less obvious directions.
Meta-reflection.
Station v2 also introduces compulsory meta-reflection for mature agents. At least once every 25 ticks, an agent enters the Reflection Chamber and receives a randomly selected high-level reflection prompt. The prompt typically asks GPT-5.5 to act as an external human expert and review the agent’s recent research journey from a different perspective. During this reflection, GPT-5.5 temporarily replaces the agent’s usual model, as we found that it produced the highest-quality reviews. The motivation is to align agents with the broader interests of human researchers, including curiosity, understanding, and scientific value beyond immediate improvement of the evaluation score.
Stagnation protocol.
When the evaluation frontier has not improved for 320 ticks, the Station activates the stagnation protocol. The protocol sends a system message to every mature agent. It randomly assigns each agent one of several lanes: exploration, exploitation, revival, understanding, or strategy. Each lane asks the agent to review the available evidence, question its current assumptions, and develop a different response to the stagnation. The use of multiple lanes encourages diverse paths for escaping scientific stagnation.
Multistart.
Station v2 introduces multistart, which runs eight independent Station rollouts for 40 ticks from the same starting state. A GPT-5.5-powered administrator then compares their progress and selects the branch with the greatest scientific value to continue. Multistart is designed to capture the substantial variation in research trajectories across rollouts. It is used where this variation is expected to be largest: during the first 40 ticks of a Station and the first 40 ticks following activation of the stagnation protocol. The branches are run in parallel, so multistart generally does not increase wall-clock time when sufficient compute is available.
Appendix B Sources of the pre-AlphaEvolve literature column
The pre-AlphaEvolve literature curve of Figure 1 is a reproducible reference assembled from work predating AlphaEvolve. No single paper tabulates these finite values. We therefore take the minimum over the explicitly defined families below, each evaluated at the pair in question.
Bukh–Chao [13], Proposition 11. We use the quadratic-residue block and the recursion , with embedded in a horizontal hyperplane. Proposition 11 makes this Kakeya for every full translation . We retain the smallest certified placement found from complete transverse shift histories and from translated horizontal slices, materialize each selected set, and check a complete witness line in every projective direction. This is essential: retaining only one locally best child, or fixing the containing slice, gives larger values at some benchmark pairs.
Kopparty, Lev, Saraf and Sudan [42]. Lemma 17 gives the upper bound (its displayed strata may overlap), and the missing-digit construction of Theorem 7 has exact size , which is the classical at .
Blokhuis–Mazzocca [11]. In the problem is settled. The minimum is exactly for odd , with a matching construction.
Products. A product of Kakeya sets is Kakeya of exactly the product size, so every product of best bounds in complementary lower dimensions is admissible, with the sharp planar value above as the factor.
Lev [46]. At , the exact value and the bound , both from a computer search.
The Bukh–Chao recursion supplies the selected value at all 22 pairs with . At , the values 13 and 27 are smaller in dimensions 3 and 4, while in dimension 5 the recursive value, the missing-digit construction and all give 63. Products never attain the minimum on their own at any pair in range. Comparing the two reference curves against each other, AlphaEvolve is below the pre-AlphaEvolve literature at 18 pairs and the pre-AlphaEvolve literature is below AlphaEvolve at 5, namely both pairs in , and the three larger primes in .
Table 3 reports all 25 benchmark pairs. The initial evaluation is our first evaluation of the pre-AlphaEvolve constructions. The final pre-AlphaEvolve literature column takes the minimum over the families described above after incorporating the extended placement search within the Bukh–Chao recursion. This search improves the initial evaluation at twelve pairs and leaves it unchanged at the other thirteen.
| Initial Evaluation | Pre-AlphaEvolve Literature | AlphaEvolve | Station | |
| 13 | 13 | 15 | 13 | |
| 53 | 53 | 53 | 53 | |
| 129 | 129 | 128 | 128 | |
| 440 | 440 | 438 | 437 | |
| 699 | 698 | 697 | 697 | |
| 2,034 | 2,034 | 2,031 | 2,030 | |
| 3,509 | 3,509 | 3,505 | 3,504 | |
| 6,837 | 6,837 | 6,833 | 6,833 | |
| 8,295 | 8,295 | 8,290 | 8,288 | |
| 13,867 | 13,866 | 13,861 | 13,861 | |
| 18,709 | 18,708 | 18,701 | 18,701 | |
| 21,504 | 21,504 | 21,495 | 21,495 | |
| 27,899 | 27,899 | 27,892 | 27,889 | |
| 39,687 | 39,686 | 39,677 | 39,677 | |
| 27 | 27 | 31 | 27 | |
| 164 | 163 | 162 | 161 | |
| 529 | 528 | 527 | 527 | |
| 2,689 | 2,689 | 2,687 | 2,684 | |
| 4,973 | 4,972 | 4,966 | 4,962 | |
| 13,524 | 13,521 | 13,514 | 13,509 | |
| 20,593 | 20,586 | 20,583 | 20,579 | |
| 63 | 63 | 63 | 53 | |
| 503 | 497 | 510 | 490 | |
| 2,145 | 2,142 | 2,187 | 2,135 | |
| 16,348 | 16,307 | 16,427 | 16,288 |
These twelve changes do not alter the comparison tally: the Station remains strictly smaller than the better reference at 14 pairs, tied at 11 and worse at none.
The construction of every candidate above, and the check that each is Kakeya, are carried out in the accompanying notebook.