An AI agent just solved 49 of 50 unpublished, research-level theoretical physics problems, at less than $3 a problem. The problems ranged from fluid flow over a plate to quantum materials to black holes. No human physicist could work across all of those fields in one afternoon. The machines did.
My first reaction was two words: virtual polymaths. A polymath is someone at home in many fields, like Leonardo da Vinci or Ben Franklin. After reading the paper, I think the label is half right. This post covers which half, and why the half that’s wrong matters for anyone who works in security.
What TheoPhysEval is
TheoPhysEval is a test of 50 physics problems, built by 35 researchers. They come from places like the International Centre for Theoretical Sciences, Harvard, Duke, KU Leuven, and Google DeepMind. The paper went up on arXiv on October 9, 2026.
The key word is unpublished. Every answer was kept off the internet, so an AI could not have seen it during training. The authors estimate a world expert could solve a typical problem in about a day. Someone from a neighboring specialty would take much longer.
Most problems were made by taking a calculation from a contributor’s recent paper and changing some of the numbers. That gives a brand-new answer nobody can guess. To solve one, you have to find the original paper, rebuild the calculation, and adapt it. The authors call these “turn-the-crank” problems. The results are new, sometimes even publishable, but they aren’t breakthroughs.
The 50 problems cover 11 fields:
| Field | Problems |
|---|---|
| Quantum field theory | 7 |
| Gravitation and cosmology | 6 |
| Classical physics | 5 |
| Stochastic processes and biological physics | 5 |
| Quantum phases of matter | 5 |
| Scattering amplitudes | 5 |
| Holography and entanglement | 5 |
| Electronic structure theory | 4 |
| Nonequilibrium statistical mechanics | 3 |
| Lattice theories | 3 |
| Particle physics | 2 |
The agents had no human hints. In the “realistic” track they could search the web and write their own scripts for numerical and symbolic math. In the “isolated” track the internet was turned off.
The results
Five agents were tested. Each one is a model paired with a “harness,” the software that lets it run commands and use tools.
| Agent (model + harness) | Solved, web on | Solved, web off | Cost, all 50 (web on) |
|---|---|---|---|
| GPT-6-Astra + Codex | 49 | 46 | $116.28 |
| Opus 5.5 + Claude Code | 46 | 46 | $240.43 |
| Gemini 3.8 Flash + Antigravity | 34 | 31 | $95.01 |
| DeepSeek-V4.1-Flash + Codex | 27 | 25 | $37.42 |
| GLM 5.3 + Claude Code | 21 | 26 | $180 (plan, both tracks) |
The best result came to about $2.33 a problem. For comparison, that’s cheaper than a cup of coffee in Honolulu.
One detail stopped me cold. At first, the graders marked several of GPT-6-Astra’s answers wrong. When they went back with the experts, they found small mistakes in the human answer keys. After the experts fixed their own solutions, the AI’s answers were correct.
The authors also flag their own limits. Each agent ran the full test only once, and AI answers vary from run to run. So treat a gap of two or three problems as noise, not a ranking.
Where the polymath label breaks
The breadth is real. The authors write that the agents’ range is wider than any single person’s. Their calculating ability now matches world experts, even in narrow specialties.
But a true polymath does more than know many subjects. They see links between fields, and they choose which questions are worth asking. On that point, the paper raises three warning signs.
1. A lot of it is memory. Turning off the internet barely changed the scores. Some problems use formulas a human would have to look up, yet the agents still solved them. The authors conclude the models have memorized large parts of the physics research literature. Knowing a library by heart is impressive, but it isn’t the same as original thought.
2. No agent solved the backwards problem. One problem gave a curve and asked for the equation that produced it. Every agent failed. Real research often looks like this: messy data in front of you, and you have to work out what’s behind it. Forward calculation is the agents’ strength. Working backwards from evidence is still hard for them.
3. They lack “taste.” The authors say agents don’t seem to share physicists’ sense of what makes a problem interesting. When asked to invent new problems, the agents fixated on details a human would find dull. They also struggled to change direction when a problem proved too hard, which good researchers do constantly.
Physics has an old saying: “shut up and calculate.” The authors end with a twist on it. Now that machines can calculate, humans may need to “stop and think.”
A more accurate name might be virtual polymath calculators. They have huge breadth of technique, but they still need a human to point them at something worth solving.
Why this matters for security
This is a physics paper, but three parts of it map directly onto security work.
Containment is just good practice. In the no-internet track, every agent ran inside a Docker container with networking switched off. Afterward, the authors checked the agents’ activity logs and found no sign that any agent escaped the sandbox. That’s the same habit I use in my own cyber range: isolate anything you don’t fully trust, then review what it did. If physicists are sandboxing their AI agents, the rest of us should be too.
Tools without judgment are a risk. An agent that can run scripts and search the web is powerful. Black Hat Python teaches the same lesson from the attacker’s side: once code can act, what matters is who points it and where. In a SOC, an AI could run every tool in the Kali toolbox. Knowing which alert matters, and why, is still the analyst’s job, because that’s “taste” again.
Cost is part of the design. Each task ran on a small, fixed cloud machine (2 vCPUs, 4 GB of memory) with an 8-hour cutoff. Costs ranged from about $37 to $240 for the full set. Anyone studying for AWS certifications will recognize the pattern: you set memory, timeouts, and budgets up front, or the bill decides for you. Agent workloads are billed by the token, so a talkative agent costs more for the same answer.
The takeaway
AI agents are now virtual polymaths at calculating. They aren’t polymaths at deciding what to calculate. The authors themselves say human experts still need to set the direction, and that agents can’t simply be handed a major open problem to solve alone.
That’s encouraging for anyone building skills right now, me included. The routine work is getting cheaper. Judgment, curiosity, and knowing which question to ask are becoming the scarce part. Those happen to be the things you build by doing the work yourself, in public, one lab at a time.
If you want to see the problems yourself, all 50 are in the paper’s appendix. They’re also on Hugging Face, along with the full test setup.