Physical-Engineering’s Last Exam
Testing AI agents on real-world engineering.
85 expert-reviewed tasks across calculations, drawings, and design reviews.
DatasetEngineering disciplines in practice
- Design a structural frame to carry gravity and lateral loads. Structural engineering, 10 tasks
- Design a site layout for buildings, grading, and utility routes. Civil engineering, 18 tasks
- Design a switchboard and feeder layout for the connected loads. Electrical engineering, 27 tasks
- Design a cooling loop to reject the facility's heat load. Mechanical engineering, 30 tasks
Skip the drawings
Design a structural frame to carry gravity and lateral loads.
Structural engineering, 10 tasksA research initiative by scientists at Google
Academic Partners
Leaderboard
Full leaderboardFrontier models evaluated on the private set, ranked with their native harness.
How grading works
Each task runs in an isolated container with the project's drawings, spreadsheets, specs, and local services.
Automated scripts check quantitative outputs, while an LLM judge grades open-ended engineering work against an expert rubric.
- No internet access
- 6 hours per task
- Three runs per harness and model
Engineering disciplines
All 85 tasks
Contribute a task
Share a realistic engineering problem with its starting files and grader. Accepted tasks join the benchmark with author credit.
See what makes a good taskCompetition
Build an agent harness, pair it with the model of your choice, and compete on the hidden test set.
Go to the competitionCitation
@misc{han2026ple,
title = {{PLE}: Physical-Engineering's Last Exam},
author = {Rujun Han and Yufan Zhuang and Arijit Ray and Han Yu and
Zoey CuiZhu and George Lee and Yifei Ming and Binz Roy and
Varun Sakalkar and Vishy Tirumalashetty and Burak Gokturk and
Tomas Pfister and Chen-Yu Lee},
year = {2026},
eprint = {2610.30537},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2610.30537}
}