ICML论文提出模型"欺骗"或"抵抗关闭"的标准验证方法
“When a model looks like it "deceives" or "resists shutdown," how do we know it isn't role-play, ins...
两位学者提出的标准能帮助区分模型行为，对部署和监管很重要。
Xin Chen和Cynthia Chen的ICML论文指出，当模型表现出"欺骗"或"抵抗关闭"时，现有证据无法区分这是角色扮演、指令遵循还是任务完成压力。论文提出在四个阶段建立共享标准来加强证据。
“When a model looks like it "deceives" or "resists shutdown," how do we know it isn't role-play, ins...
“When a model looks like it "deceives" or "resists shutdown," how do we know it isn't role-play, instruction-following, or just task-completion pressure? Often the current evidence can't yet tell them apart, and those claims are starting to inform deployment and regulation.” - @XinCynthiaChen Excellent thread, with important recommendations for researchers in how they talk about their work. Xin Chen, Cynthia @XinCynthiaChen Excited that our ICML position paper has been accepted as an Oral 🎉! When a model looks like it "deceives" or "resists shutdown," how do we know it isn't role-play, instruction-following, or just task-completion pressure? Often the current evidence can't yet tell them apart, and those claims are starting to inform deployment and regulation. We map where the evidence gets thin across four stages and propose a shared standard to strengthen it. Here's the map. � arxiv.org/abs/2606.07612 5P 🔗 View Quoted Tweet 💬 3 🔄 2 ❤️ 4 👀 2411 📊 3 ⚡
