← Blog

Explainer19 min read26 sources

Physicists ask AI companies to release results openly: the high-energy physics statement

On 13 November 2026, 40 physicists asked that major AI-company results in their field be released openly and in full. What it says, and what of Yang–Mills is public.

Share

On 13 November 2026, a group of theoretical high-energy physicists published “A Statement on Sharing Research Results Produced by AI Companies in Theoretical High-Energy Physics” at hep-open-release.org1. Its central sentence says that “major research results produced by AI companies in our field should be released openly and in full”, so that “all researchers, wherever they work, can engage with new results on an equal footing”. Forty physicists signed first, among them Edward Witten, Juan Maldacena, Nathan Seiberg, Cumrun Vafa, Sheldon Glashow and David Tong. When we read the page again on 17 November 2026, it listed 261 further signatories.

Much is left out of the statement. It names no company and no result, sets no deadline, and does not define “major” or “in full”. It binds no one. The link to OpenAI's Yang–Mills work comes from the Proofs and Prompts newsfeed, which wrote on 16 November that the statement is “partially in response to OpenAI's release of their progress on Yang-Mills which was initially shared only with a selected group of experts”.2 The statement's own page does not say what prompted it. We found no public reply from OpenAI or any other AI company as of 17 November 2026.

Below: the statement and its signatories, the Yang–Mills release it is read against, a timeline, earlier calls from mathematics, the web and arXiv precedents, what “openly and in full” could mean, and what open release has made possible since 10 November. Our earlier posts covered the shepherding committee's first statement and the group's first assessment. Every source was checked on 17 November 2026.

What does the high-energy physics statement say?#

It has four short paragraphs and a list of names.1 Theoretical high-energy physics is the part of physics that describes elementary particles and forces with mathematical theories such as quantum field theory and string theory.

The first paragraph names the change: “breakthrough results are now being produced by internal large language models at AI companies.” It limits its own scope: “Our focus is not on the research itself, but on how its results are shared.” The companies, it says, “must navigate many fields, each with its own culture and expectations”, and the statement gives the view of one field.

The second paragraph offers two precedents from that field's history, the World Wide Web and arXiv. It says arXiv “made research preprints available online to everyone at once”, which “largely removed the inequity that existed when preprints circulated by mail and researchers at well-connected institutions often saw new results well before everyone else.”

The third paragraph is the request: “We, the undersigned theoretical high-energy physicists, believe that major research results produced by AI companies in our field should be released openly and in full.” It gives two reasons. Open release puts all researchers “on an equal footing”, and “It also helps the community absorb them more quickly, by allowing researchers to scrutinize and discuss them together.”

The last paragraph is one sentence to the companies. It begins “We recognize that AI companies are still working out how best to share their results” and says that releasing them openly “would honor and continue our field's storied tradition of free and open exchange.”

Who signed the statement, and how many have signed?#

The 40 initial signatories work at institutions in ten countries: 21 in the United States, five in India, three each in the United Kingdom and the Netherlands, and the rest in Israel, Canada, Brazil, France, Japan and China.1 Besides the six named above, they include Ofer Aharony, Shiraz Minwalla, Robert Myers, Michael Peskin, Stephen Shenker, Tadashi Takayanagi and Spenta Wadia.

A second list, “Further signatories”, grows as names are approved. Our first copy of the page, downloaded at 00:35 Turkey time on 17 November (21:35 UTC on 16 November), listed 261 further names from institutions in 29 countries, the largest groups in the United States (63), India (36), the United Kingdom (34) and Italy (22). A second download at 00:56 Turkey time showed the same list.

The page has a form for adding a name and explains the process: “After you submit the form, we will email you a link to confirm your signature. The organizers then review each confirmed signature before it appears on this page.” It prefers an institutional email address, “because it helps us confirm your affiliation.” The page does not name the organizers or say who drafted the text.

One overlap is worth knowing. Edward Witten is also one of the nine members of the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), hosted at the Institute for Advanced Study, which made its own call for release in October.11

What does the statement leave out?#

The text names no AI company, no model and no specific result.1 It sets no deadline and asks for no venue, licence or format. It does not say what makes a result “major”, or what a release “in full” must contain, and it does not mention formal proofs, prompts or model versions. It proposes no sanction and speaks only for “our field”. What the signatories would count as a full release is left to the reader; the table further down gives one reading, marked as ours.

What is the statement responding to?#

The statement does not say. Proofs and Prompts, a site that collects “recent breaking news and updates from the worlds of Mathematics and AI”, listed it on 16 November as an “Open letter” with its own reading, quoted above.2 The newsfeed adds that sharing open letters “does not constitute an endorsement.”

The release it refers to began with OpenAI's announcement of 26 October 20263, which set out three results and three release strategies. For the physics result, OpenAI's outside experts advised: “form an external group of collaborators and defer to them on the release of materials.” The announcement went on: “For the physics collaboration, the group has decided to hold off on releasing the AI-produced artifacts and instead work on writing it up themselves.” OpenAI named the trade-off itself: “choosing initial shepherds necessarily shapes who gets involved first and creates asymmetry of information. We welcome feedback on how to broaden participation and make this process more useful.”

The physics claim is broader than Yang–Mills. OpenAI describes “a mathematically rigorous existence proof of certain interacting quantum field theories containing gauge fields and fermions in four dimensions, including Quantum Chromodynamics (QCD) and a unified completion of the Standard Model of particle physics”, with pure Yang–Mills theory as a subproblem.3 It says its models produced the Yang–Mills proof on 9 September and the QCD proof on 18 September, and that it finished the Lean formalization of Yang–Mills existence on 16 October. Lean is a proof assistant: software that checks every step of a proof written in its formal language.

The group that took custody, chaired by David Gross with Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov, set out its approach in a first statement dated October 2026 (the PDF was created on 27 October): “Before we release it, we would like to be convinced that it is correct.”4 It would first check the definitions, then release the Lean proof, and “When we are satisfied with the quality and correctness of the proofs, we will provide our assessment together with the full 1147-page LLM proofs.” It called the arrangement “an experiment in how the academic community can deal with the rapid development of new knowledge by AI.”

On 3 November the group wrote that the Lean proof would come “hopefully next week”, and that “In all cases we will announce any release of LLM produced material a full day in advance.”5 On 9 November it announced publication for the next day.6 The Lean code and the group's 35-page Assessment 1 went public on the night of 10 November, US time; the first commit in the group's results repository is timed 03:37 UTC on 11 November.8

What was released, and what is still unreleased?#

Public since then: OpenAI's Lean source as submitted on 28 October, in its own archive10 and copied into the results repository, with the group's checks, dated evidence and rerun instructions.8 The assessment judges that the submission “proves that a relativistic quantum field theory satisfying the Wightman axioms can be constructed as a scaling limit of four-dimensional Wilson lattice gauge theory for any simple compact gauge group. It does not prove that the infinite-volume limit is unique or that there is a mass gap.”7 Our 11 November post explains the checks.

Still not public on 17 November:

  • The informal manuscript. The assessment says it “does not draw on the separate informal proof manuscript.”7 The 1,147 pages are in none of the group's three GitHub repositories, nor in OpenAI's openai/math repository, last changed on 30 October. The KITP page that carries the group's statements had no update after 9 November.26 Science, on 11 November, called it “the still-unreleased proof”.23
  • The second submission. The results repository lists “OpenAI, 4 November 2026” as “work in progress”.8
  • QCD and the unified theory. The group's overview has three result categories, YM, QCD and chiralSU5 (chiral SU(5) gauge theory), and says: “Links for QCD and chiralSU5 will be added when materials are released.”9
  • How the result was found. We found no published prompts, transcripts, attempt counts or model name.

The group's own terms changed on 15 November. Its overview and the results repository gained a sentence: “Our normal arrangement is that submitters retain ownership of their work and remain free to post or publish it elsewhere, independently of our review.” The group adds that it “will consider other arrangements for custody and release on request”, naming OpenAI's 26 October arrangement as one.9 The same day, the results repository listed a new submission, “Alpöge and Sussman, 15 November 2026”, also as work in progress.8 Neither change mentions the physicists' statement.

Timeline: from OpenAI's announcement to the statement#

Date (2026)Event
9 SeptemberOpenAI's models produce a natural-language Yang–Mills existence proof, according to OpenAI3
16 OctoberOpenAI finishes the Lean formalization, by its account; Gross is asked to take custody of the proof4
25 OctoberAGMAI recommends that AI companies stop designating mathematicians to receive output first11
26 OctoberOpenAI's announcement: the physics group will hold off on releasing the AI-produced artifacts.3 AGMAI calls for release “without delay”11
27 OctoberThe committee's first statement: Lean proof after validation, the 1,147 pages later with an assessment4
28 OctoberOpenAI submits its Lean code to the group7
3 NovemberSecond statement: release “hopefully next week”, announced a day ahead5
4 NovemberSecond OpenAI submission, still listed as work in progress on 17 November8
9 NovemberThird statement announces publication for 10 November6
10 November, night (US)Lean code and Assessment 1 published; first repository commit at 03:37 UTC on 11 November8
10–11 NovemberTwo public projects start turning the Lean code into readable mathematics2021
13 NovemberThe high-energy physics statement is published1
15 NovemberThe Yang–Mills group states its “normal arrangement” for submitters and lists the Alpöge–Sussman submission9
16 NovemberProofs and Prompts reads the statement as partly a response to the Yang–Mills release2
Table 1 From OpenAI's Yang–Mills announcement to the physicists' statement, as dated in the sources below. OpenAI's September and October dates are its own statements.

What had mathematicians already asked for?#

Calls from mathematics came earlier, and some say more precisely what a release should contain.

Who, whenWhat it asksScope
Leiden Declaration, 2 June 2026, endorsed by the International Mathematical Union15Warns that evaluation suffers when results come through press releases or blog posts without “disclosure of information necessary for scientific evaluation”Mathematics
Terence Tao, Mathstodon, 8 September12Points to “the refusal of AI companies to disclose their negative results, or reveal the process towards obtaining their solutions”Mathematics
AGMAI, 25 and 26 October11No designated first readers; immediate release of the Yang–Mills outputs; model name, prompts and costs with any releaseMathematics, “not necessarily” other sciences
Harvard CMSA summit report, late October14“There is value in AI labs sharing the step-by-step reasoning processes behind their solved problems whenever possible.”Mathematics
High-energy physics statement, 13 November1Release “openly and in full”Theoretical high-energy physics
Table 2 Calls for the open release of AI-produced results, in order of date.

AGMAI made two statements around OpenAI's announcement.11 On 25 October it recommended “that AI companies do not adopt this practice for future releases”, meaning the practice of designating particular mathematicians to receive output before anyone else. Its first reason: “In cases where designated mathematicians see material before everyone else, they get a head start on problems others may have worked on for years.” It added that this “should in no way be construed as criticism of the mathematicians who have taken on such a role”. On 26 October it wrote: “We call on OpenAI to make public without delay the natural-language and Lean outputs behind its announced existence proofs for Yang–Mills theory, quantum chromodynamics, and a completion of the Standard Model.”

AGMAI also says what a release should include: output posted “on scholarly repositories that are not controlled by the posting AI lab, have citable identifiers, and record modifications”, with “the name of the model, prompts, chains of thought, number of attempts, compute used, and costs”. And it marks its limit: “We recall that our recommendations are intended for mathematics, and not necessarily for other areas of science.” The physicists' statement now gives a view for one area of physics, in that field's own name. AGMAI's page listed no statement after 26 October when we checked.

Terence Tao's concern ran in two directions. On 8 September he wrote about companies not disclosing their process.12 On 6 October he wrote about researchers: “promising open directions are now being withheld from the public”, for fear that their own work would be scooped.13 We covered his argument in Is AI using up math's open problems?

The Harvard CMSA report, from a summit of 32 mathematicians on 22 and 23 October, explains why the reasoning matters as well as the result: “This transparency allows the mathematical community to identify the underlying patterns and problem-solving strategies used by machines.”14 Our post on the report covers the rest.

What do the web and arXiv precedents refer to?#

Both are documented by the institutions involved.

The web. CERN's history says: “Tim Berners-Lee, a British scientist, invented the World Wide Web (WWW) in 1989, while working at CERN. The web was originally conceived and developed to meet the demand for automated information-sharing between scientists in universities and institutes around the world.” And: “On 30 April 1993, CERN put the World Wide Web software in the public domain.”16 CERN's licensing page names what was released, “the three components of Web software (the basic line-mode client, the basic server and the library of common code)”, and quotes the 1993 statement: “CERN relinquishes all intellectual property rights to this code, both source and binary and permission is given to anyone to use, duplicate, modify and distribute it.” For version 3 of its server software, announced on 15 November 1994, CERN kept the copyright under its own open-source licence, so that the software stayed free and could not be turned into proprietary software by others.17

arXiv. arXiv's about page says it “was founded by Paul Ginsparg in 1991”.18 Its archive of early news has two entries for that year: in June, “Announcement of proposed hep-th archive sent out for community comment”, and on 14 August, “e-mail interface for hep-th put on-line”. An entry of May 1992 lists hep-th as “high energy physics formal”, with 1,075 subscribers.19 The statement's account of mailed preprints and unequal early access is its own; we did not find an arXiv page that describes it in those words.

What could “openly and in full” mean for an AI-produced result?#

The statement does not say, so the table below is our reading. It lists the parts a full release could contain, drawing on what AGMAI names and on what the Yang–Mills group itself published, and checks each against the Yang–Mills case on 17 November.

Part of a releaseWhat it lets a reader doOpenAI's Yang–Mills work on 17 November 2026
The claimed result, stated exactlyKnow what is claimedPublic. The Lean theorem YangMills.jointExistence, the group's stronger combined target, and a statement in words in the assessment7
The natural-language proofRead, teach and check the argumentNot public. The 1,147-page manuscript is unreleased; community drafts written from the Lean code exist21
The formal proofRebuild it and check what it statesPublic. 9,477 Lean files at a fixed commit, Apache-2.0 licence, with rerun instructions10
Prompts, transcripts or a method descriptionSee how the result was found, and what failedNot public. OpenAI gives discovery dates, not prompts, attempts or compute3
Model and version, as reportedKnow which system produced itOpenAI writes “our internal models”;3 the committee writes “OpenAI's LLM”.4 No name or version
Who checked whatWeigh the checkPublic. Rebuild, axiom audit, a check against a fixed reference statement, the group's own Lean checks, and human review of the definitions7
Related results and later versionsSee the whole claimNot public. The 4 November submission and the QCD and chiral SU(5) materials9
Table 3 What a full release of an AI-produced result could contain, in our reading, and what we found public for OpenAI's Yang–Mills work on 17 November 2026.

The assessment is candid about the limits of its checks. It reports “extensive (but not exhaustive) human review of the Lean definitions” and says: “We do not claim to fully understand the proof yet.”7 It was prepared by Douglas, Jaffe, Ron Nissim and Fred Rajasekaran and approved by the whole group, which says it used OpenAI's Astra “to help translate the Lean to informal mathematics” and Anthropic's Fable “to give feedback on this document”. Those are the group's tools, not the model that produced the proof.

A Lean check covers what the code states. Whether those statements mean what physicists mean by Yang–Mills theory is a question people answer, which is why the natural-language proof and the human reading of definitions matter as much as the code.

What has open release made possible since 10 November?#

Once the Lean code was public, anyone could read it. Within hours, two public projects on GitHub were at work on it, both using AI agents.

Humanise Yang–Mills, set up by Martin Hairer on 10 November, imported the Lean source at 06:49 UTC on 11 November. It “coordinates a crowd-sourced effort to turn the announced Lean formalisation of four-dimensional Yang–Mills theory into a human-readable proof”, aiming at “a research monograph giving detailed proofs of the argument”.20 Its README gives the status as “source imported; preliminary dependency mapping under way” and says that “A full build, axiom audit, and line-by-line mathematical audit are still outstanding.” Contributors send AI agents to take small tasks; on 17 November GitHub's search counted 361 merged pull requests. The README states the limit of the code: “Passing Lean checks does not establish that an English argument accurately describes the formal theorem; that correspondence requires review.” Hairer is also a member of AGMAI.11

Yang–Mills informalization, maintained by Scott Armstrong since 11 November in collaboration with Hairer's project, describes the source as “about one million lines of Lean in 9,477 files. No informal proof accompanies it.”21 It maps the proof into eight parts and thirty-two units and offers a third-draft manuscript of 278 pages, dated 13 November, “written entirely by AI agents from the Lean code, and is still being revised.” Its warning: “The outline is a rough draft written by machines. Its purpose is to be checked, corrected and rewritten by people.” It also traces where the proof's ideas come from, component by component, against Balaban's programme and other constructive work.

Neither project says the proof is correct, and an AI-written manuscript is itself a claim that needs review. What open release changed is who can take part. Armstrong told Scientific American on 11 November: “I'm happy the shepherds are releasing the Lean proof, finally, so we can see what the result is. And I hope, in the future, the AI labs will share their results widely with the community.” He added: “This review can be done more efficiently and quickly in the open.”22 Science reported the same day that Ethan Sussman, who works with Anthropic's Levent Alpöge on a separate Yang–Mills effort, wants research open from the outset: “I want to invite everyone to a collaboration,” he said.23

Has OpenAI or any other AI company replied?#

We found no public reply as of 17 November 2026. We searched news coverage and OpenAI's news feed after 13 November. On 15 November OpenAI launched a science blog, and its first guest post, by Nathaniel Craig, a visiting scientist at OpenAI, described a candidate lattice definition of the Standard Model developed with GPT‑6 Astra.2425 Neither post mentions the statement.

What does this mean for a research record?#

Whatever form a full release takes, a few facts make an AI-produced result usable later, and anyone citing or building on one can write them down now:

  • Who did what. OpenAI's models produced the proof, according to OpenAI; the Yang–Mills group checked the Lean code; public projects are writing it up with AI agents.
  • Which model, as reported. Here only “internal models”, with no name or version. Note the absence too.
  • Which version. The Lean source at commit 3456e5e of the 28 October submission, Assessment 1 of 10 November, the informalization draft of 13 November. Cite a commit or a dated file; READMEs change.
  • What was released, when, and what was not. The informal manuscript, the 4 November submission and the QCD materials were not public on 17 November.

Papex's page on Yang–Mills existence and mass gap lists the problem as open and records OpenAI's claim with the model as OpenAI states it and the Lean repository at that commit. A record of who released what, and when, does not show that the mathematics is right. It does keep the physicists' question answerable later: who could see a result, and from when.

Sources#

  1. A Statement on Sharing Research Results Produced by AI Companies in Theoretical High-Energy Physics (hep-open-release.org, 13 November 2026), checked 17 November 2026 a b c d e f
  2. Newsfeed: Open letter by Theoretical High-Energy physicists calling for open release of results produced by AI (Proofs and Prompts, 16 November 2026), checked 17 November 2026 a b c
  3. Working with academic communities to build scientific understanding (OpenAI, 26 October 2026), checked 17 November 2026 through the Internet Archive copy of 15 November 2026, because openai.com blocked automated reading a b c d e f
  4. Proof by OpenAI of the Existence of Quantum Yang-Mills theory: Statement by the shepherding committee (dated October 2026; PDF created 27 October 2026), checked 17 November 2026 a b c d
  5. Proof by OpenAI of the existence of Quantum Yang-Mills Theory: 2nd Statement by the Shepherding Committee (3 November 2026), checked 17 November 2026 a b
  6. Proof by OpenAI of the existence of Quantum Yang-Mills Theory: 3rd Statement by the Shepherding Committee (9 November 2026), checked 17 November 2026 a b
  7. Validation of the OpenAI Yang–Mills construction: Assessment 1 (Yang–Mills group, 10 November 2026, 35 pages), checked 17 November 2026 a b c d e f
  8. yang-mills-group/YM: README and commit history, last changed 15 November 2026 (GitHub), checked 17 November 2026 a b c d e f
  9. Yang–Mills group organization overview, last changed 15 November 2026 (GitHub), checked 17 November 2026 a b c d
  10. yang-mills-group/OpenAI-YM-20261028: OpenAI's Lean submission of 28 October 2026 (GitHub), checked 17 November 2026 a b
  11. AGMAI statements of 25 and 26 October 2026, and About AGMAI (Advisory Group on Mathematics and Artificial Intelligence), checked 17 November 2026 a b c d e f
  12. Terence Tao, thread on open problems as drinking water beside an ocean (Mathstodon, 8 September 2026), checked 17 November 2026 a b
  13. Terence Tao, thread on “Math 1.0” and “Math 2.0” (Mathstodon, 6 October 2026), checked 17 November 2026 ↩
  14. Report of Summit on Math Research in the Age of AI (Harvard CMSA, summit of 22–23 October 2026), checked 17 November 2026 a b
  15. Leiden Declaration on Artificial Intelligence and Mathematics (2 June 2026), checked 17 November 2026 ↩
  16. The birth of the Web (CERN), checked 17 November 2026 ↩
  17. Tim Smith and François Flückiger, Licensing the Web (CERN), checked 17 November 2026 ↩
  18. About arXiv (arXiv), checked 17 November 2026 ↩
  19. arXiv news archive from 1991–1994 (arXiv), checked 17 November 2026 ↩
  20. hairer/humanise-yang-mills: README (GitHub, started 10 November 2026), checked 17 November 2026 a b
  21. scottnarmstrong/yang-mills-informalization: README (GitHub, started 11 November 2026), checked 17 November 2026 a b c
  22. Joseph Howlett, Scientists confirm OpenAI's blockbuster quantum physics result (Scientific American, 11 November 2026, updated 12 November), checked 17 November 2026 ↩
  23. Jacklin Kwan, AI rivals tackle one of physics' deepest mathematical challenges (Science, 11 November 2026), checked 17 November 2026 through the Internet Archive copy of 15 November 2026, because science.org blocked automated reading a b
  24. Our new blog for science (OpenAI, 15 November 2026), checked 17 November 2026 through the Internet Archive copy of 16 November 2026 ↩
  25. Nathaniel Craig, A step toward defining the Standard Model on a lattice (OpenAI, 15 November 2026), checked 17 November 2026 through the Internet Archive copy of 16 November 2026 ↩
  26. Statement by the Yang-Mills Working Group, with the 3 and 9 November updates (KITP), checked 17 November 2026 ↩

Share

Keep reading

All posts