CoPersona uses peer graphs to fix sparse LLM personalization

Real-world LLM personalization suffers from sparse, biased user histories—most users leave only a handful of interactions, and even heavy users' logs miss key preferences. When a request shifts to an under-supported facet (e.g., a user's niche movie genre preference), personalization fails. CoPersona, presented by Yangtian Zhang and colleagues, solves this by borrowing signals from behaviorally similar peers. It decomposes each user's interaction history into multiple facet-level representations (e.g., genre, price range, style), then builds a multiplex persona graph where nodes are users and edges represent facet-level similarity. This avoids the bias of mixing all signals in a single global space.

At inference time, CoPersona uses a dual-branch architecture: a non-parametric branch retrieves top-k peer personas for the active facet, while a parametric graph reasoning branch propagates information across the multiplex graph. The two branches are combined to generate a personalized response. Experiments across domains like movie recommendations, product reviews, and task-oriented dialogues, tested on models from 1B to 70B parameters, show CoPersona consistently outperforms strong baselines like fine-tuning and prompt-based personalization. The paper is accepted at KDD '26, highlighting its practical value for scalable, robust personalization.