Can LLMs Replace Survey Respondents?
中文摘要
研究探讨如何通过“遗忘”技术解决大语言模型在模拟问卷调查时出现的“模式崩溃”问题,以提高合成回复的多样性。
English Summary
Researchers explore using unlearning to fix mode collapse in LLM-generated survey responses, ensuring more diverse and realistic synthetic data.
原文节选
How unlearning fixes mode collapse in synthetic survey replies