AI is moving from mathematical proof toward broader scientific theory
Artificial intelligence systems are beginning to show measurable progress on theoretical questions in fields beyond mathematics, according to research discussed by MIT researcher Neil Thompson. The analysis suggests that the capability gains now unsettling parts of mathematics could extend to probability, statistics, economics, high-energy theory, gravitation, cosmology and quantum physics.
The finding is not that AI has replaced scientists or that it can independently run a modern research program. The study focused on a narrower but consequential task: solving theoretical problems whose answers were later established in research papers. In those domains, the researchers found a sharp improvement in the strongest models’ success rates between mid-2024 and mid-2026.
That trajectory matters because theoretical work often supplies the language, models and predictions on which experimental science depends. If researchers can reliably use AI to explore portions of that work, the effect may be felt not only in individual calculations but in the pace and organization of scientific inquiry.
From a small share of problems to more than a third
Thompson’s study, conducted with Gabriel Manso and Claudio Mayrink Verdun, tested a series of OpenAI models on problems drawn from research papers posted between December 2025 and May 2026. The group examined models including o1, o3, GPT-5, GPT-5.1, GPT-5.4 and GPT-5.5.
Across the theoretical areas included in the assessment, the best systems moved from correctly solving about 2% of the problems in mid-2024 to about 35% by mid-2026. That is still far from universal competence, and it does not mean every correct-looking answer is dependable. But the scale of the increase is notable: a tool that succeeds rarely may be a curiosity, while one that can make substantial contributions across a benchmark can become part of ordinary research practice.
The researchers designed the evaluation to reduce a central concern in testing language models: whether a system is simply reproducing material encountered during training. The problems came from papers posted before the relevant solutions appeared in later publications, and the published solutions served as the ground truth. Internet access was blocked during testing, so the systems could not retrieve an answer from the web.
That setup does not eliminate every limitation of benchmark-based evaluation. It does, however, aim to distinguish problem-solving from straightforward lookup or memorization. The study’s result is therefore evidence that the tested models could generate solutions to at least some questions for which the answer was not available in their accessible training material.
A bounded result, not a general scientific substitute
The scope of the work is important. The team concentrated on fields where models could address problems without needing to run simulations or conduct physical experiments. Much of science cannot be reduced to symbolic reasoning alone. Laboratory work, instrument design, field observation, data collection, simulation and replication remain essential parts of discovering how the world works.
Even within theory-heavy disciplines, a correct solution is not the same as a useful research contribution. Scientists need to judge assumptions, choose worthwhile questions, interpret results and determine whether a proposed argument holds up under scrutiny. Those are social as well as technical processes, involving peer review, debate and cumulative evidence.
Still, the areas chosen for the study are not peripheral. Probability and statistics shape inference throughout science and industry. Economic theory influences policy and market analysis. High-energy physics, cosmology and quantum physics are fields in which abstract reasoning can be central to advancing a question before it reaches a detector, telescope or laboratory apparatus.
If AI becomes capable of handling more of the routine or intermediate work in these areas, the immediate effect could be uneven. Researchers may use it to generate candidate approaches, check derivations, explore alternative formulations or accelerate early-stage analysis. The value of those uses will depend on verification. A fast path toward a wrong conclusion can waste time just as effectively as a slow calculation.
The mathematics precedent is shaping expectations
The new assessment arrives as mathematics is already confronting highly visible demonstrations of AI capability. OpenAI recently announced that an unreleased model had solved a longstanding grand challenge involving the Navier-Stokes equations, which describe fluid motion. That episode drew controversy, underscoring that extraordinary claims in mathematics and AI require careful independent examination.
For Thompson, the broader implication is that theoretical disciplines may be approaching a period in which AI’s role changes quickly. He expects many theory questions to be transformed within a few years by what the systems can do. The study’s trend line is the basis for that expectation, rather than proof that any particular field will be fully automated.
The most plausible near-term future is collaborative rather than solitary: scientists setting agendas and standards, while AI systems take on a growing share of problem exploration. That could widen access to sophisticated analytical tools, but it could also make the ability to validate machine-generated work more valuable. Research institutions will have to decide how to credit contributions, train students and preserve the independent expertise needed to catch subtle mistakes.
What to watch next
- Whether performance gains persist on newly created problems rather than only on retrospective evaluations.
- How often model-generated solutions remain correct after expert review and formal verification.
- Whether systems can contribute usefully when simulations, experiments or specialized data are required.
- How scientific groups adapt authorship, training and quality-control practices as AI becomes a regular research tool.
The study does not establish that AI will become the default method for solving scientific theory problems. It does show why that possibility is now being taken seriously. In less than two years, the strongest tested systems went from solving a small fraction of selected questions to solving more than a third. For disciplines built around difficult reasoning, that is a shift worth watching closely.
This article is based on reporting by Wired. Read the original article.
Originally published on wired.com
