99Science · AI & science

SCIENCE UNDER AUTOMATION

AI in Theoretical physics: capabilities and evidence

Removing the laboratory does not make the whole field easy to automate. Choosing a tractable question and finding a useful new description remain distinct from solving a specified problem.

What has AI demonstrated?

Current benchmark is under review

Test
CritPt independent leaderboard: leading listed result 32.9%; not used as validated evidence.
Comparison
Research-level answer grading; validity review unresolved.
System and version
CritPt independent leaderboard · 2026-10-20
Evidence checked
2026-10-20

Excluded from passing quantitative validity gates while under review; no replacement for corrected scores.

What does the field require?

Solve a specified research physics problem

Develop a new physical theory

Research tasks

Solve a specified research problem

CritPt is the closer mapping. FrontierScience pools multiple scientific disciplines.

Develop a new theoretical description

Novelty and explanatory value need assessment beyond a correct final answer.

Connect a theory to observations

The timing and quality of observations can dominate this stage.

Can AI replace the whole workflow?

The results above measure specific tasks. The field assessment also requires completion of the connected workflow, including the requirements listed here. A field-wide replacement percentage has not been established.

Read the assessment method

Sources and related evidence

Compare related fields

Computational genomics · Experimental molecular biology · Field ecology