Google 联合 UCLA、CMU 发布 GenXR-Bench：AI 编程工具做 XR 应用表现不佳
Google 和 UCLA、CMU 造了个 XR 编程测试基准，112 个任务一测，最好的模型也就刚过一半，坑都在 3D 空间定位上。
Google、UCLA 和 CMU 的论文构建了 GenXR-Bench，包含 112 个专家编写的 XR 开发任务，涵盖功能开发、Bug 修复和完整应用构建。测试中 Claude Opus 5 以 50.9% 的成绩领先，但没有模型在 SDK 功能任务上超过 29.2%。失败的运行中 64.4% 出现坐标系错误，44.9% 出现启动竞态条件。该基准的模拟测试与真实头显下的盲测 XR 专家判断吻合率达 96.5%。
New Google + UCLA + CMU paper finds that AI coding tools often fail to build virtual and mixed reality apps because they lose track of where things are in 3D space.
Claude Opus 5 led at 50.9%, and no model topped 29.2% on SDK feature tasks. Coordinate-frame mistakes appeared in 64.4% of failed runs, and startup race conditions in 44.9%.
XR apps run in 3D space and react to head motion, hands, and room sensing. Normal coding benchmarks use unit tests or browser clicks, which cannot check whether a virtual button follows your palm.
They built GenXR-Bench, 112 expert-written tasks covering XR feature work, bug fixes, and full app builds. Its tests simulate hands and rooms, and matched blinded XR experts using real headsets 96.5% of the time.
If you use agents for XR apps, test the result with simulated hand and room input, since agent-written unit tests often miss runtime bugs.
