AI agents solve 49 of 50 unpublished physics problems in DeepMind-backed study

The AI model GPT-6-Astra successfully solved 49 of 50 unpublished theoretical physics problems in a new preprint study titled 'TheoPhysEval: Assessing AI agents at autonomous theoretical physics problem solving.' Published on arXiv on October 9, 2026, the research represents a significant advancement in AI's capabilities for scientific reasoning.

Benchmark details

The benchmark contained 50 research-level physics problems spanning 11 subfields including fluid dynamics and quantum field theory. These problems had never been published, making them more challenging for AI models that might have memorized existing solutions.

Physicists from institutions such as ICTS and Google DeepMind led the research, with 35 contributors from various research institutions. The researchers tested off-the-shelf models rather than custom-built systems, using criteria designed to approximate expert theoretical physicist skills.

Performance results

Testing occurred in two tracks: 'realistic' with web access and 'isolated' without. GPT-6-Astra scored 49/50 with web access (costing $116.28) but dropped to 46/50 without it. Another model, Opus 5.5, scored 46/50 in both tracks.

Research implications

The study revealed that the AI models' work led to corrections of some human-written reference solutions, suggesting they could identify errors in existing research. The $116.28 cost for solving 50 problems is significantly lower than the time investment required by human experts.

The researchers noted important limitations: the study is a preprint未经正式同行评审, and the benchmark covers only 50 problems. Additionally, solving well-posed problems differs from the creative work of formulating new research questions, which remains a uniquely human capability in theoretical physics.