跳到正文
arXiv · cs.AI · Alexandre Cristov\~ao Maiorano · 6 小时前模型评分44

合成人设能预测真实受众反应吗?基于 Upworthy 标题排序的通用读者基线研究

Do Synthetic Personas Predict Real Audience Response? A Headline-Ranking Study with a Generic-Reader Baseline

导读

研究用 Upworthy 标题对比固定十人设面板与通用读者基线,在 Gemini 3.1 Flash-Lite 和 GPT-4.1 上各取 190 个标题包、每臂每变体 30 次打分。

分类置信度

中 · 待人工复核

  • 只有单一媒体来源,需要人工复核。

来源级别:媒体与个人

事件类型:研究发布

证据摘录

  • “Do demographic personas improve an LLM's prediction of audience engagement?”
  • “These intervals condition on the selected cohort and recorded generations and do not adjust for dependence between related experiments.”
  • “The results concern this panel, cohort, and protocol; they are not a held-out evaluation or evidence of universal persona harm.”
  • “Provider request protocols differ across models, and no causal mechanism is isolated.”

来源:arXiv · cs.AI · arxiv.org