The authors recruited 87 participants who built personal memory profiles from their actual ChatGPT queries.
Personalized Memory Doesn't Always Improve User Preference for LLM Responses
Dong et al. find that users frequently prefer responses without their personal history, motivating a new alignment objective for memory-augmented models.
Top University
Yijiang River Dong · Zexue He · Cheng Yu · Tiancheng Hu · Michael J Ryan · Jan Batzner · +6 more
University of Cambridge · Stanford University · Technical University of Munich · Harvard University · University of Texas at Austin
Research Digest··2 min read
Dong et al.
§