视频选题榜 2026-09-14

• https://arxiv.org/abs/2609.10226 · Φ-Bench · Can Large Language Models Engineer the Infrastructure That Powers Them? · StepFun AI 出品·Opus 5 36.5% 不及格·AI 跑不过自己·评测方法学双栖预备触发·harness 偏差+E2EO reward hacking 风险 · round=R1