In short
New research suggests frontier AI models are rapidly improving on theory-heavy science problems, with performance jumping from about 2% to 35% in two years. The advance could speed up discovery in fields like physics and economics, but it also raises serious questions about verification and human oversight.
- Frontier models are increasingly capable of solving theory-heavy scientific problems.
- A new study tested models on papers published after their training cutoff to avoid data leakage.
- Performance reportedly rose from about 2% in mid-2024 to about 35% by mid-2026.
- Fields like probability, economics, cosmology and quantum physics may be next in line for disruption.
- Experts say human researchers will still be needed to validate results and generate ideas.
Artificial intelligence is starting to reshape theory-heavy science far beyond mathematics, with new research suggesting frontier models can now solve a growing share of problems in fields such as statistics, economics, cosmology, gravitation and quantum physics. The shift matters because it could accelerate scientific discovery while also raising fresh questions about verification, credit and the role of human researchers.
That is the central takeaway from a study discussed by Neil Thompson at MIT, where he argued that AI systems are rapidly improving on theoretical problems that once seemed out of reach. The findings suggest the recent upheaval in mathematics may be the start of a broader transformation across science.
Thompson, who studies the economics of computing and AI, presented research indicating that the best models moved from solving roughly 2% of benchmarked theoretical problems in these disciplines in mid-2024 to about 35% by mid-2026. If that pace continues, whole classes of scientific reasoning could soon be assisted — or in some cases partially automated — by AI.
But the story is not simply one of machines replacing scientists. The emerging picture is more complicated: AI may become powerful at generating candidate solutions, while humans remain essential for asking the right questions, checking the math and deciding whether the answers actually hold up.
What the new research shows
The study highlighted by Thompson was designed to test whether AI systems could reason through genuinely new theoretical questions, rather than merely repeat patterns learned from older material. That distinction is important, because many benchmark tests can be inflated by data contamination or by training examples that overlap with test problems.
To avoid that problem, Thompson and two colleagues — PhD student Gabriel Manso and research scientist Claudio Mayrink Verdun — drew problems from research papers posted online between December 2025 and May 2026. They then evaluated several OpenAI models, including o1, o3, GPT-5, GPT-5.1, GPT-5.4 and GPT-5.5, to see whether the systems could derive the answers on their own.
The researchers took extra precautions. The models were cut off from the internet, preventing them from looking up solutions. And because the papers used for testing were published after the training cutoff for the models, the answers should not have been part of the systems’ preexisting knowledge.
In other words, the models had to reason, not recall.
How did the researchers measure progress?
They measured progress by comparing model performance across a set of theory-focused tasks in which the answer could be checked against published solutions. The aim was not to test experimental science, where success may depend on laboratory equipment, simulations or fieldwork, but to focus on areas where formal reasoning is central.
That choice narrowed the scope, but it also sharpened the signal. Fields that rely heavily on abstract derivation and formal structure appear to be moving into AI’s sweet spot faster than areas that depend on messy empirical work.
| Field or area | Why it matters | AI relevance |
|---|---|---|
| Mathematics | Foundation for formal reasoning and proof | Already seeing major disruption |
| Probability and statistics theory | Core tools for inference and modeling | Strong fit for symbolic and logical reasoning |
| Economic theory | Uses abstract models and formal argument | Increasingly vulnerable to AI assistance |
| High-energy theory | Relies on mathematical structure and derivation | Likely to be transformed by stronger models |
| Gravitation and cosmology | Theory-heavy areas with formal equations | Suitable for benchmarked reasoning tasks |
| Quantum physics | Abstract and mathematically dense | Potentially next in line for AI breakthroughs |
Why these fields may be next
These disciplines share a key property: many of their hardest problems can be stated precisely on paper. That makes them accessible to systems that are good at logic, algebra, pattern completion and multi-step derivation.
According to the study’s framing, AI does not need to run a physical experiment to make progress in these areas. It may only need to reason through formal steps, test internal consistency and produce a correct derivation. That is a task frontier models are getting much better at.
Thompson’s broader point is that scientific progress often follows a chain reaction. A theoretical breakthrough can later unlock experimental advances, new algorithms or fresh applications. If AI accelerates the first step in that chain, the downstream effects could be substantial.
The implications extend beyond academic prestige. Theory work often feeds into engineering, finance, computing and applied sciences. Faster progress in foundational fields could therefore ripple outward into technologies and industries far from the original research topic.
How is this different from previous AI progress?
The speed of improvement stands out. In a little over two years, the benchmark success rate described by Thompson’s team appears to have climbed from near-novelty levels to a point where AI is solving more than one-third of the tested problems.
That kind of jump echoes what has already happened in mathematics. For many researchers, the mathematics shift was a warning sign that theory-heavy sciences were next. The new work suggests that warning was justified.
OpenAI’s recent activity has intensified that sense of momentum. The company recently said an unreleased model had solved a famous open problem involving Navier-Stokes equations, which govern fluid motion. The announcement caused a stir not just because of the result, but because it triggered debate over priority and credit after another group accused OpenAI of moving in on work they believed they were close to publishing.
The broader effect has been a scramble across the math and AI communities. OpenAI and others have disclosed more solutions to longstanding problems, feeding enthusiasm on one side and alarm on the other.
“Within just a few years, we expect a lot of theory questions to be totally transformed in terms of AI’s ability to do them,” Thompson said, underscoring how quickly he believes the field is changing.
What scientists worry about
Not everyone sees these advances as an unqualified good. Some mathematicians have warned that if AI can solve a problem without helping humans understand the path to the solution, then science may lose some of its explanatory value. In mathematics, the route to an answer is often as important as the answer itself.
That concern applies even more broadly to science. A result that cannot be independently verified is of limited use, especially in fields where a subtle error can invalidate an entire line of reasoning.
Manso, one of Thompson’s co-authors, said the promise of AI in science should not be confused with certainty. He stressed that researchers will still be needed to generate ideas, connect concepts and assess whether a model’s answer is actually trustworthy.
Manso argued that humans will still be essential for coming up with promising directions and for checking whether an AI’s output really holds together, especially on difficult problems where mistakes can be hard to spot.
He also warned that models can be overconfident. A system may state that it has found the right answer, yet the full derivation can reveal weak links, unjustified leaps or hidden errors. That is one reason why verification tools may become as important as the models themselves.
Why verification could become the bottleneck
If AI becomes good at generating candidate solutions faster than humans can inspect them, the challenge could shift from discovery to validation. That would create a new kind of scientific bottleneck.
In that scenario, the limiting factor would no longer be idea generation alone. It would be the ability to confirm results rigorously, reproduce them and understand their implications. That is especially true in theoretical science, where a subtle mistake can masquerade as a convincing proof.
The issue has already surfaced in OpenAI’s own work. The company reportedly withdrew some solutions from a recent dataset after publication, a reminder that even advanced systems can produce answers that do not survive closer scrutiny.
How AI could change the scientific workflow
The most immediate change may not be total automation. Instead, AI could become a high-speed collaborator that drafts proofs, explores hypotheses and suggests candidate derivations for human experts to review.
That would alter the rhythm of research. Scientists might spend less time grinding through routine algebra or proof search and more time on framing questions, checking edge cases and synthesizing results across disciplines.
This could be especially valuable in fields where researchers are overwhelmed by the complexity of the problems and the volume of possible approaches. An AI system that rapidly narrows the field of plausible solutions may give human experts a much better starting point.
At the same time, there is a real risk that overreliance on model outputs could create shallow progress. If researchers accept answers too readily, errors may spread quickly. If they use models only as black boxes, they may gain speed but lose understanding.
The most likely outcome, at least in the near term, is a hybrid model of science: AI proposes, humans dispose.
Who is building AI for science?
AI-focused science has become a serious area of investment and experimentation. Some prominent AI researchers have launched new labs explicitly dedicated to scientific discovery, while major model developers have increased their emphasis on research-oriented products and tooling.
Those efforts range from theory assistance to systems designed to help with laboratory work, automated experimentation and hypothesis generation. The long-term goal is not merely to answer homework-style questions, but to create systems that can contribute to real scientific progress.
Even so, the practical hurdles are large. Theory may be the easiest part of science for AI to absorb, but science is not only theory. Many breakthroughs depend on measurement, instrumentation, physical intuition and judgment that cannot be reduced to symbols on a page.
That is why many researchers see current progress as a beginning rather than an endpoint. The jump into theory-heavy science may be the first stage of a much wider AI transition, but it is unlikely to be the last obstacle.
What happens next?
If the trend continues, more scientific subfields will likely become testbeds for frontier AI models. Benchmarks will get tougher, datasets will expand and researchers will be forced to ask which tasks are best done by machines and which still depend on human insight.
There will also be pressure on universities, journals and funding agencies to adapt. They may need new standards for evaluating AI-assisted work, clearer rules for disclosure and better tools for checking whether a result is independently sound.
At the same time, the scientific community may benefit from the speedup. Problems that once took months of expert labor could be explored in days. Ideas that were too tedious to pursue manually may suddenly become practical.
Still, the central tension will remain: the better AI gets at solving theory problems, the more urgent it becomes to distinguish between a result that is merely plausible and one that is truly understood.
Timeline of the AI-for-science shift
The latest findings sit inside a much faster-moving sequence of breakthroughs across math and science. The following timeline captures the key turning points discussed in the study and surrounding debate.
| When | What happened | Why it mattered |
|---|---|---|
| Mid-2024 | Top AI models solved about 2% of selected theory problems | AI performance on formal science looked limited |
| Late 2025 to spring 2026 | Researchers assembled a fresh benchmark from newly published papers | Created a test set beyond the models’ training data |
| Mid-2026 | Performance rose to about 35% on the same style of problems | Suggested a dramatic acceleration in reasoning ability |
| Recent months | OpenAI said an unreleased model solved a major Navier-Stokes challenge | Raised the profile of AI in advanced mathematics |
| Now | Researchers are looking beyond math to other theory-heavy sciences | Indicates the next wave of disruption may already be underway |
What this means for science as a whole
The biggest lesson from Thompson’s work is that AI’s scientific impact may be broader than many expected. What began as concern about math is now looking like the opening move in a wider transformation of theoretical research.
That does not mean human scientists are being sidelined. Instead, their role may change. The value of expertise may shift toward problem selection, conceptual framing, validation and synthesis — the parts of science that remain hardest to automate.
For now, the evidence points to a future in which AI is not just a tool for writing, summarizing or coding, but a genuine participant in the production of new scientific knowledge. The promise is enormous. So is the need for caution.
As Thompson’s findings suggest, the next frontier of AI may not be language at all. It may be the logic of science itself.
This article is based on a talk and study discussed by MIT researcher Neil Thompson and his collaborators, with additional context on recent AI and mathematics developments.
Frequently asked questions
How is AI changing science beyond mathematics?
AI is beginning to solve formal, theory-heavy problems in other disciplines such as statistics, economics, cosmology and quantum physics. The main change is that models are getting better at multi-step reasoning on abstract questions that resemble mathematical proofs.
Why did researchers use newly published papers in the benchmark?
They used newly published papers to make sure the models could not have seen the answers during training. That made the test more rigorous by forcing the systems to reason through the problems rather than memorizing prior solutions.
Does this mean AI will replace scientists?
No, not in the near term. AI may take over more routine derivations and candidate solution generation, but scientists are still needed to frame questions, verify answers and judge whether results are mathematically or scientifically sound.
Which fields appear most vulnerable to AI disruption next?
The study points to probability, statistics theory, economic theory, high-energy physics, gravitation and cosmology, and quantum physics. These fields rely heavily on formal reasoning, which makes them more accessible to current frontier models.
What is the biggest risk of using AI in theoretical science?
The biggest risk is that models may produce convincing but incorrect answers. Because verification can be difficult in advanced theory, scientists could waste time or accept flawed results unless there are strong review and checking systems.
