DeepMind Paper: AI Agents Solve 49 of 50 Unpublished Physics Problems

1 reports archived0 independent sourcesupdated 10/25/2026, 18:42:05

Synthesis & Latest Updates

A new study by Google DeepMind and top universities evaluates AI agents on real-world research tasks. GPT-6-Astra with web search solved 49 out of 50 unpublished theoretical physics problems at under $3 each. Notably, models without search still solved 46 problems, suggesting significant internalization of physics literature.

LATEST/A new study by Google DeepMind and top universities evaluates AI agents on real-world research tasks. GPT-6-Astra with web search solved 49 out of 50 unpublished theoretical physics problems at under $3 each. Notably, models without search still solved 46 problems, suggesting significant internalization of physics literature.

Hourly heat curve

94 fully observed hours

Now

0.0

Peak

20.020:00

24h change

–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first

Rohan Paul (@rohanpaul_ai)T2Oct 25, 18:42·88 pts
DeepMind Paper: AI Agents Solve 49 of 50 Unpublished Physics Problems
- GPT-6-Astra solved 49 of 50 unpublished physics problems (equivalent to a day's work for an expert) with web search.
- Accuracy dropped minimally without search (46/50), implying extensive memorization of physics literature.
- Cost per problem remained under $3, highlighting high economic viability for research assistance.
View Related Reports & Analysis Original Source