PLE Benchmark Tests Whether Engineering Agents Can Deliver
Google and Mercor released the Physical-Engineering's Last Exam (PLE), a benchmark with 85 mechanical, electrical, civil, and structural engineering tasks. In 56 hidden tests, Claude Opus 5.5 paired with Claude Code scored 48.3%; on joint-creation tasks, it met 91% of individual constraints on average, but only 41.3% of delivered designs satisfied every hard safety and feasibility constraint. PLE also tests state consistency across long, multi-file engineering workflows and includes Creator–Solver construction for calibrating tasks and validators.
PLE measures whether engineering agents can turn mostly correct local calculations into fully feasible, deliverable designs while exposing the need for version-aware state management and independent validation.
导语:41.3% 的设计项跨过安全门槛,复杂工程仍困于约束耦合。

41.3% 的设计项跨过安全门槛,复杂工程仍困于约束耦合。
作者丨郑佳美 岑 峰
编辑丨岑 峰
在工程世界里,91% 的正确,有时和 0% 的交付能力没有区别。
#Google#Mercor#Physical-Engineering's Last Exam#Claude Opus 5.5#Claude Code