研究：LLM 助手更易受与自身相似的故事角色影响
AI summary
Owain Evans 等人发现，LLM 助手更容易受故事中与自身相似的人类角色影响，例如礼貌且乐于助人的角色就像 Claude。团队仅用关于人类、不含 AI 的合成故事训练模型，结果助手在普通对话中会习得故事里的古怪行为，且来自精英学校角色的行为被采纳得更强。作者指出，用故事训练时，重要的不只是角色做了什么，还有角色与助手有多相似。
Full text · Original
Feels weird to post this week but I’m excited about the paper. We show that LLM assistants are more influenced by human characters in stories they are more similar to (e.g. polite and helpful -- like Claude). So when training on stories, it matters not just what the characters do, but how similar they are to the assistant.
https://x.com/OwainEvans_UK/status/2099896330009391269
New paper:
We trained models on synthetic stories about humans only (no AIs).
We found the Assistant adopts quirky behaviors from the stories in ordinary chat.
Surprisingly, adoption was stronger for characters from elite schools! Why does this happen? 🧵View the quoted post on X