无提示词时LLM在想什么？模型家族主题偏好研究
原标题：What do LLMs think when you don't tell them what to think about?
AI 摘要
Together AI 的研究团队通过使用极简、无主题的提示词，研究了大型语言模型在无明确指令时的自然生成行为。他们发现不同模型家族表现出显著不同的主题偏好，如 GPT-OSS 偏向编程和数学，Llama 偏向文学，DeepSeek 常生成宗教内容，Qwen 则倾向于生成选择题。这些差异具有系统性，并揭示了模型在无约束条件下的退化文本模式，可能带来安全和隐私风险。该研究为模型审计和行为监控提供了新视角。
正文节选
Our interactions with large language models (LLMs) are dominated by task- or topic-specific questions, such as “solve this coding problem” or “what is democracy?” This framing strongly shapes how LLMs generate responses, constraining the range of behaviors and knowledge that can be observed. We study the behavior of LLMs using minimal, topic-neutral prompts. Despite the absence of explicit task or topic specification, LLMs generate diverse content; however, each model family exhibits distinct to