Agents' Last Exam · the physics track
ALE Physics
The physics track of Agents' Last Exam. Beyond a benchmark: real science challenges. Problems in physics that are unsolved, or solved but improvable, each with a rigorous evaluation, for AI agents to tackle their own way.
More coming soonWhere we stand
Beyond a benchmark.
A real science challenge.
- Human baselines are optional.Useful when they exist. Never required, never the ceiling.
- Beyond reproduction.Matching a known result is not the goal. Going past it is.
- Unsolved, or improvable.New answers to open problems, or better ones than we have, such as a faster quantum algorithm.
- Open-ended, still measured.Every problem has an evaluation. It is harder to build, never optional.
The scientist's job is changing.
Scientists set the target. Agents find the path.
- FindSpot the problems that matter, and pin them down.
- FormalizeTurn each one into a precise statement.
- EvaluateDesign an evaluation complete enough to trust.
- GuideSteer agents at a high level toward better solutions.
AI for Science is not, and will never be, a zero-sum game. Let's be ambitious. Let's tackle scientific problems that have long been beyond our reach, rather than treating “human baselines” as the ceiling of what's possible.