在麻省理工学院（MIT）一场演讲中，Neil Thompson 提出前沿人工智慧模型正迅速颠覆多个理论科学领域。由 Neil Thompson、Gabriel Manso 与 Claudio Mayrink Verdun 共同进行的研究指出，在 2024 年中至 2026 年中的两年间，顶尖人工智慧模型正确解答机率论、统计理论、经济理论、高能理论、重力与宇宙学以及量子物理学等理论问题的能力，成功率已从约 2% 跃升至约 35%，展现出惊人的进展速度且毫无减缓迹象。
该研究评估了 OpenAI 旗下的 o1、o3、GPT-5、GPT-5.1、GPT-5.4 及 GPT-5.5 等模型，测试题目选自 2025 年 12 月至 2026 年 5 月间发表于线上的学术论文。为确保严谨评估，所有论文均发布于模型训练截止日期之后，并在测试时彻底阻断网路连线，迫使模型独立推导未曾见过的解答。此项研究聚焦于无需实体实验或数值模拟的纯理论领域，显示人工智慧在高度依赖数学与理论推演的学门中所引发的变革正在持续扩散。
近期人工智慧在理论界的冲击显著加剧，OpenAI 宣布解决长久未解的 Navier-Stokes 方程问题，随后又公布了数百个长年数学难题的解答。然而，这波浪潮引发了学界对验证机制与过度依赖工具的疑虑。尽管各大实验室正在扩展人工智慧在科学研究中的应用，学者们强调人类专家的监督仍然不可或缺，特别是在产出核心观点、跨领域连结，以及严格审查人工智慧可能产生错误推导的复杂解题过程方面。
In an MIT talk, Neil Thompson presented evidence that frontier artificial intelligence models are rapidly transforming several theoretical fields. A study conducted by Neil Thompson alongside Gabriel Manso and Claudio Mayrink Verdun revealed that between mid-2024 and mid-2026, the success rate of premier artificial intelligence models in solving theoretical problems across probability, statistical theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics escalated from approximately 2 percent to about 35 percent, displaying a sharp upward trajectory without slowing down.
The evaluation analyzed OpenAI models, including o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5, against theoretical problems extracted from research papers posted between December 2025 and May 2026. Crucially, these benchmark problems appeared after the training data cutoffs of the tested architectures, and internet access was strictly blocked during evaluation to ensure that solutions were derived independently rather than retrieved. The investigation focused specifically on theory-heavy domains that bypass empirical laboratory experiments or computer simulations, underscoring how mathematical automation is expanding.
This shift follows major milestones, notably OpenAI announcing a solution to the Navier-Stokes equations followed by hundreds of additional solutions to longstanding mathematical problems. While these advances demonstrate profound capabilities, they have also ignited debates regarding verification and the loss of problem-solving processes. As major laboratories accelerate scientific AI integration, researchers stress that human domain expertise remains vital to formulate key concepts, synthesize discoveries, and rigorously audit model outputs against subtle errors in complex derivations.