OpenAI publishes AI research evidence in mathematics, physics, and computer science
OpenAI says it has produced three research results in mathematics, computer science, and physics, then formalized them with the help of external experts. The company is testing a new dissemination mode, balancing transparency, verification, and scientific credit.
OpenAI claims to have generated and then formalized three research results in mathematics, theoretical computer science, and physics, announced on October 26, 2026. One of them purports to refute the Exponential Time Hypothesis, a central conjecture in algorithmic complexity; another deals with the local smoothing conjecture, and the third with an existence proof for certain quantum field theories, including QCD. This topic matters because the company is not just selling more talkative models: it is also experimenting with how to turn AI outputs into verifiable scientific results, a point that touches on methodology, credit, and trust.
A scientific dissemination designed as a test, not a classic announcement
OpenAI says it considers these AI-generated manuscripts as raw research data, not as completed science. For each result, the company asked external experts to act as shepherds, in other words pilots responsible for organizing reading, verification, and contextualization. This setup was meant to choose between several ways of publishing: direct dissemination of the AI text, forming a group of external collaborators, or delayed publication of materials when researchers judge them ready.
This approach is based on four principles announced by OpenAI: transparency, no non-disclosure agreements, no academic credit claimed by the company, and independence of participants. In short, groups can discuss, share their ongoing work, and publish their conclusions. OpenAI also specifies that it does not speak on behalf of the researchers, and vice versa. A rather rare framework for cutting-edge research. Almost a social experiment with equations.
Why the proof on the Exponential Time Hypothesis is making waves
The first result, in the field of theoretical computer science, concerns the Exponential Time Hypothesis, a hypothesis about the difficulty of certain Boolean satisfiability problems, closely related to P versus NP. OpenAI says it produced a refutation, then verified in Lean, a formal proof assistant that allows rigorous verification of mathematical proofs. The natural language output dates from October 16, 2026, the formalization from October 17, 2026.
This hypothesis has served as a basis for hundreds of papers establishing conditional hardness results in areas such as graph theory, computational biology, databases, or algorithmic geometry. If the refutation is confirmed, it does not instantly invalidate these works; it mainly requires revisiting a portion of the literature based on this hypothesis. The announced verification group is to be led by Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams, with the goal of publicly examining source files and associated proofs.
In physics, OpenAI targets QCD and the Standard Model
The third result concerns mathematical physics. According to OpenAI, its models produced an existence proof for certain four-dimensional quantum field theories, with gauge fields and fermions, including quantum chromodynamics, or QCD, as well as a unified completion of the Standard Model. The company specifies that this work relies on Balaban’s approach in constructive quantum field theory and includes, as a subproblem, an existence proof of pure Yang-Mills theory.
The timeline matters here as well: OpenAI indicates having obtained a natural language proof for Yang-Mills on September 9, 2026, then for QCD on September 18, 2026, before awaiting the complete Lean formalization, finished on October 16, 2026. The working group is to be led by David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov. For this part, the company chose not to immediately publish the AI-produced artifacts and to let the group draft the reference version themselves.
What this method changes for research
The core of the matter is not only the solidity of the three results; it is the way to integrate them into the scientific circuit. OpenAI itself acknowledges that the selection of the first facilitators creates an information asymmetry and influences subsequent exchanges. In this context, the company says it wants to receive feedback on what works, what blocks, and what needs improvement to broaden participation.
The implications are quite direct for laboratories, publishers, and funders: who verifies, who understands, who cites, who reuses, and at what pace? OpenAI says it wants to fund the independent review of the three results, notably through a $5 million contribution to the Simons Foundation to support workshops, conferences, and special programs. The company also hopes these collaborations will help better understand the capabilities, but also the limits, of efficient computation. In short, AI has not yet replaced the researcher; it already forces research to rethink its assembly line.
Key points
- OpenAI announces three research results on October 26, 2026.
- The computer science proof concerns the Exponential Time Hypothesis.
- The physics result concerns QCD and the Standard Model.
- Lean is used to formalize and verify the proofs.
- Russell Impagliazzo and David Gross are among the leaders.
- OpenAI mentions $5 million for the Simons Foundation.
By the numbers
- 3 results — mathematics, computer science, physics, announced on October 26, 2026 by OpenAI.
- 4 principles — transparency, no NDA, no claimed credit, independence.
- $5 million — announced contribution to the Simons Foundation.
- 2 dates in October 2026 — October 16 and 17 for the computer science proof; October 16 for the Yang-Mills formalization.