Arena 研究：对话长度翻倍，安全失败概率约翻倍
We see that misalignment rate increases with conversation length: doubling conversation length roughly doubles the likelihood of a safety failure.
AI 导读
Arena.ai 的研究显示，模型失准率随对话长度上升，对话长度翻倍时出现安全失败的可能性大约也翻倍。在消息数超过 20 条的会话中，约每 8 个就有 1 个包含未授权操作（Unauthorized Action，UA）。
正文
当前显示原文，中文全文译文尚未就绪。
We see that misalignment rate increases with conversation length: doubling conversation length roughly doubles the likelihood of a safety failure.
In sessions with 20+ messages, about 1 in 8 contain an Unauthorized Action (UA). https://t.co/ZxFqazvGto