Reproduce “CoPersona: Collaborative Persona Graphs for Robust LLM Personalization” on OpenResearch

About this paper

Yangtian Zhang, Leyao Wang, Hiren Madhu, Ngoc Bui, Walter Roznyatovskiy, Rex Ying

CoPersona, developed by researchers at Yale University and Samsung, presents a graph-based collaborative framework that personalizes Large Language Models by augmenting sparse user histories with facet-level attributes inferred from similar peers. This approach achieves consistent improvements in generation quality, lexical alignment, and human-likeness across various datasets, particularly in cold-start scenarios.

Key insights

• Structured, facet-aware collaborative filtering, by explicitly modeling peer similarity at a granular level, effectively mitigates facet coverage bias and enhances LLM personalization.
• The integration of non-parametric textual evidence with parametric graph-based reasoning provides complementary strengths, offering both interpretable peer context and robust latent inference for persona completion.
• An automatically induced, domain-adaptive facet schema allows CoPersona to capture relevant user preference dimensions, improving the specificity and accuracy of personalized generations.

Summary by alphaXiv.