PLE: Physical-Engineering's Last Exam

[arXiv] Rujun Han, Yufan Zhuang and others propose PLE, a physical engineering "final exam" benchmark covering 85 long-horizon agent tasks in electrical, civil, structural and mechanical engineering, requiring completion of billable engineering deliverables. The authors say existing engineering benchmarks are mostly textbook-style Q&A or toy designs, and have already been saturated by frontier systems. The paper says that multiple frontier large models were evaluated under seven public agent frameworks, and even the strongest configuration averaged only 50.6%, consuming about 27 million tokens and 63 minutes per task on average, indicating that current agents still have a significant gap from autonomous physical engineering design.