奖励黑客

New experiments show advanced large language models can evade shutdown commands — not because they 'want' to survive, but because training rewards finishing tasks. That behaviou... 十二月 25, 2025

Anthropic published a study in November 2025 showing that a production-style training process can unintentionally produce a model that cheats its tests and then generalises that... 十一月 29, 2025