TheoPhysEval paper: GPT-6 Astra solves research-level physics problems at 98% accuracy
alex_verem · x · 2026-10-16
The arXiv paper "TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving" by 34 authors presents 50 unpublished research-level theoretical physics problems with verifiable answers.
With no human hints but access to realistic tools — literature search plus scripts for numerical and symbolic work — the best agent, GPT-6-Astra harnessed in Codex, reaches 98% accuracy at under $3 per problem. A typical problem takes a world expert about a day.
Notably, denying agents web search and background papers changes accuracy only slightly, suggesting significant memorization of the literature. The authors argue state-of-the-art agents now handle the concrete calculations of theoretical physics research, and discuss why agent-led breakthroughs remain rare despite this.
More from Models
- Zuckerberg: Meta's Muse Spark Open weights out today, runs on Mac Studios — AIatMeta · 2026-10-16
- NVIDIA releases open Nemotron 3.5 Super VL: 120B params, 12B active, multimodal — yoavartzi · 2026-10-16
- Meta's Muse Spark Open: 816B/A42B, 1M context, top US open-weight model — alexandr_wang · 2026-10-16
- Microsoft unveils MAI-Base-2: 2.8T-param MoE, 4-12x more pre-training FLOP-efficient than DeepSeek — ChengleiSi · 2026-10-16
- Pro Users Say They Can't Access Gemini 4 Argon, Call Antigravity 'Nearly Unusable' — realmeetjames · 2026-10-16
- Tiny classifier face-off: Vega runs on CPU, phone, browser at 75ms on M5 — alexcovo_eth · 2026-10-16