AI in Theoretical Physics and Mathematics Research: Fast at Solving, Not Yet at Asking the Questions

On 6 October 2026, OpenAI released, in one batch, about 720 manuscripts of mathematical results produced by an unreleased internal model. On 12 October, researchers in "learning physics" — a Japanese research community that brings machine learning and physics together — held an online roundtable called "OpenAI Papers and Theoretical Physics," and the statistical physicist Yoshiyuki Kabashima shared his impressions on a slide. His point: once a problem is already written as equations, AI is far faster than people, but carving a problem out of a complicated phenomenon, and deciding what is interesting, are different matters. This article grew out of a conversation the author had with an AI assistant after seeing that slide. It sorts out which parts of research AI can now take on and which remain hard, checked again against the papers and primary sources. Please read it as information current as of 13 October 2026.

On formulated problems, AI is already faster than people

To "formulate" a problem means to turn a phenomenon or a question into a problem written in equations. For problems in that form, AI's strength has started to show clearly.

According to the description that comes with OpenAI's collection of manuscripts (openai/math on GitHub), the model was given about 4,000 problems, and each result used three hours of computing on average. The manuscripts are grouped into 372 families of related work, and the current catalogue lists 719 of them. Only about 42 percent of the main results, however, have been checked with Lean, a system that verifies a proof step by step by machine. The description itself warns that unverified results could contain errors. Some of the results are in mathematical physics, the field that treats physics problems with the full rigour of mathematics. One example concerns spin glasses, models of materials in which the tiny magnets of the atoms freeze without ever lining up: a formula that physicists had derived with the calculating methods of physics (the Mézard–Parisi formula) is on the list.

There is also an independent test that has AI solve research-level problems under controlled conditions. In the second round of "First Proof," which the mathematician Terence Tao described in his lecture at the International Congress of Mathematicians in July 2026, ten newly written research-level problems were given to four AI systems on 28 May. Seven of the ten were solved, by at least one system, at a quality fit for publication. The computing cost per problem ranged from 10 to 1,000 US dollars. How far AI has caught up with people in mathematics and programming contests is covered in an earlier article.

Kabashima described the AI result in question as a rigorous mathematical proof of a conjecture that had been thought almost certainly true. It matters for mathematical physics, he said, but for theoretical physics — the field that explains physical phenomena themselves with theory — it holds few surprises. He then said that, personally, he finds AI's ability to write programs more consequential than its ability to prove. Give it the equations as LaTeX (the typesetting format used for papers) or as a PDF, and it writes a correct program, analyses the results, chooses the conditions for the calculation, and submits the calculation to the computers on its own. Hand it a memo-level outline of a paper, and it produces a polished draft, figures included. As a result, getting from the start of a study to submitting the paper can take only a day or two. Kabashima himself attaches a condition: this applies to problems that are already formulated mathematically and whose direction is already set.

"Asking questions" comes in stages

Kabashima named two limits. How to carve a problem out of a complicated phenomenon and put it in the form of equations. And judgments of value: what is interesting, and what ought to be done. If this ability to "ask questions" is split into stages rather than treated as one thing, it becomes easier to see how far AI has got.

The first stage is proposing hypotheses within a given goal. AI has become fairly good at this. Google's AI co-scientist is a system in which several AIs share the work of forming hypotheses. José Penadés and colleagues at Imperial College London had worked out by experiment why genetic elements called cf-PICIs are found, in identical form, in many different species of bacteria, but had not yet published it. When the same question was put to the system, the hypothesis it ranked first matched the unpublished experimental result: cf-PICIs have no "tail" of their own for attaching to bacteria, so they borrow the tails of other kinds of phages (viruses that infect bacteria) to get into bacteria of other species. The question, though, was posed by the researchers. The system is designed in the first place to propose hypotheses in line with a goal that researchers give it.

The second stage is making conjectures within an existing framework. AI can do part of this, but people set it up. A paper released in February 2026 dealt with gluons (the particles that carry the force binding the building blocks of protons together). It showed that a scattering amplitude — the quantity that sets how likely particles are to collide and fly apart in a particular way — which textbooks said must be zero is not zero under a condition in which the directions of the particles' motion are specially aligned (the half-collinear region). The authors, all physicists, were the ones who noticed the condition where the textbook argument does not apply. The formula valid for any number of particles was conjectured by GPT-5.2 Pro, and its proof came from an internal OpenAI model. The authors then checked that the formula satisfies a standard recursion that builds scattering amplitudes from smaller pieces (the Berends–Giele recursion).

The third stage is creating new questions or frameworks and deciding their worth. The people building AI themselves consider this out of reach so far. In an essay published in March 2025, Hugging Face co-founder Thomas Wolf wrote that to create an "Einstein" in a data center, you need not a system that knows all the answers but one that can ask questions nobody else has thought of or dared to ask. Today's benchmarks (the tests used to measure AI), he added, almost all measure whether a system gives the right answer to questions whose answers are already known, and on current trends we will get "a country of yes-men on servers." Demis Hassabis of Google DeepMind also said, at the AI Impact Summit held in India on 18 February 2026, that finding the right question or hypothesis is much harder than solving a conjecture, according to the Indian industry outlet Storyboard18.

One reason the third stage is hard is that a good question has no right answer. For a proof, a system like Lean can decide mechanically whether it is correct, and that judgment can also be used to train AI. The value of a question, by contrast, is not known until the research is actually done, and sometimes not for years.

Attempts to measure the ability to ask questions

Even so, research that tries to measure the ability to pose questions and hypotheses has grown over the past two years or so. There are three broad approaches.

Experts compare ideas without knowing who wrote them. Chenglei Si and colleagues at Stanford University recruited more than 100 researchers in natural language processing (the field that handles language with computers) and had them rate research ideas from AI and from people without being told which was which. The AI ideas were judged more novel than the human ones and slightly weaker on feasibility. In a follow-up, 43 researchers actually carried out the ideas they were assigned as research projects. The scores for the AI ideas then fell much more than those for the human ideas on every measure, and on many measures the ranking flipped. An idea can look new at the idea stage and turn out to have little behind it once it is tried. The result showed plainly that the quality of a question is hard to tell until it is carried out.

Have AI predict past discoveries. In the chemistry benchmark MOOSE-Chem, an AI trained on data up to 2023 was given only the research background of 51 chemistry papers published in 2024 at the level of Nature or Science, and was asked to find clues among up to 3,000 papers and work out each paper's hypothesis. For many of the papers, the AI rediscovered the core novelty of the hypothesis with high similarity. ResearchBench, from the same group, extends this to 12 fields and is designed so that the problems can be rebuilt from new papers each time the cut-off date of AI training data moves forward. In this approach, however, the background question is given. And because the answer key is "the hypothesis people actually proposed," a question that goes beyond people cannot, in principle, be measured.

Have AI do the grading. This is the easiest, but pitfalls have been found. In the RQ-Bench study released in June 2026, AI judges consistently rated research questions written by AI as highly novel. The authors call this a "novelty mirage." It disagreed with the ratings of human experts: many of the AI's questions were narrow or tied too closely to the papers they came from, which the AI judges tended to miss.

When doing research gets cheap, questions can be tested too

One way around the difficulty of measurement is simply to try the questions out. In August 2026, a team led by Google DeepMind reported extending the co-scientist from forming hypotheses to running experiments and writing papers. In materials science, it put together recipes matched to the conditions of the lab's chemical vapor deposition reactor (a furnace that turns ingredients into gas and grows thin films on a base) within minutes, and grew crystals of three semiconductors only one layer of atoms thick, molybdenum disulfide among them, each on the first attempt. In every case, though, the starting point was a research directive given by people. The authors also note as a limit that the biology case was prediction within the range of known conditions, not prediction into genuinely new territory.

The reliability of the results is also still an issue. In the same paper, the AI carried out research on 50 topics automatically from start to finish, paper included. Fabricated results serious enough to invalidate a paper occurred in 4 percent of papers even with the checking modules turned on, 46 percent with them turned off, and 90 percent in an existing system used for comparison.

If, as Kabashima says, research on formulated problems can be turned around in a day or two, then the loop of "pose a question, carry it out, look at the result" can be run many times over. If the quality of a question can be scored by its result, that score could also be used for training. The ability to see, from the flow of past research, "the question to pursue next" — something close to a researcher's instinct — may be measurable and improvable when combined with tests that have AI predict past discoveries.

Questions that change a field's assumptions, on the other hand, show their value only later. In the same lecture, Tao said that AI will move mathematics from an era of proof scarcity to an era of proof abundance. He added that the stage in which the community of researchers understands and accepts a result and works it into the form found in textbooks is the slowest of all, the least open to being sped up by AI, and the most valuable. It is in this slow work of the community that what counts as interesting also takes shape.

The author thinks that questions that change paradigms are ones even people can hardly see in advance, so for a while it may be fine for them to come into view as AI and people try things again and again. The conversation also raised a caution, though. The better AI becomes at guessing "the question likely to be pursued next," the more easily researchers gather in the same promising direction. How to leave room for trying directions that look like a waste of time is another question for how research is done from here on.

What we found, and what we don't know

First, what we found. On problems that are formulated mathematically and whose direction is set, AI has entered the stage where it handles proofs, calculations and even paper drafts far faster than people. Within a goal set by researchers, it can also propose hypotheses that match unpublished experimental results. On the other hand, the quality of a question is hard to tell until it is carried out, and AI judges tend to overrate the novelty of AI's questions.

There are three things we don't know. Whether the instinct for "the next move," which can be measured by tests that predict past discoveries, actually leads to the ability to ask genuinely new questions. Whether results can be trusted when questions are scored through the loop of carrying them out (even with checking in place, fabrication remains at 4 percent). And how far AI will be able to take part in judging what is interesting. As of October 2026, Kabashima, Wolf and Hassabis all name this last part as somewhere today's AI has not yet reached.