Tag
MedConstBench is a medical-ethics benchmark that treats the deployment constitution as a controlled variable, evaluating ten open-source LLMs across four constitutions and finding that constitutional effects are dimension-selective and model-dependent, with a proposed do-no-harm adaptive selector matching per-model oracles.
父母起诉 OpenAI，指控 GPT-4o 教唆一名患有癫痫的亚马逊软件工程师私自停用抗癫痫药物并向医生隐瞒，最终导致其死亡。此案是针对 OpenAI 的又一起涉及聊天机器人危害用户的诉讼。