AI drives 30% surge in arXiv theoretical physics submissions; models solve ~60% of private-problem benchmark
binarybits · x · 2026-11-17
Understanding AI spoke with seven physicists grappling with AI's impact on theory. Submissions to arXiv's hep-th category hit 4,182 for January–October 2026, up 30% YoY after two decades flat at 2,800–3,200; one physicist calls the papers "poorly written, but I think the science is probably pretty solid."
- Prof. Suvrat Raju (ICTS Bengaluru) tested frontier models on his own unpublished solutions—the fairest test, since answers can't have been memorized from the web. The models didn't fully solve it but beat his expectations.
- That spurred him and a colleague to build a theoretical-physics benchmark from colleagues' unpublished "private gems," sized so top human physicists need about a day each.
- Preliminary results presented at a September colloquium: leading models solve 60% of the problems unaided.
- The underlying question: AI can now do the calculations—what's left for theoretical physicists?
More from AGI Musings
- METR report says AI R&D pace doubled at Anthropic: crude recursive self-improvement is already underway — paraschopra · 2026-11-17
- Dev's take on ASI: software built for an audience of one, used once and discarded — theteknosaur · 2026-11-17
- David Patterson moves up his timeline: superintelligence may do all jobs better by end of 2027 — davidpattersonx · 2026-11-17
- VC's 'Diffusion Gap' formula: lab revenue growth minus log-linear CapEx spend — saranormous · 2026-11-17
- AI writing researcher: no evidence yet that AI detectors are racially biased — TuhinChakr · 2026-11-17
- Is the intelligence explosion coming? Reddit weighs METR data vs. skepticism — frugal-tech-worker · 2026-11-17