Search papers, labs, and topics across Lattice.
This paper introduces CoPersona, a graph-based framework that enhances LLM personalization by addressing the challenges posed by sparse and biased user interaction histories. By decomposing user interactions into facet-level representations and employing a multiplex persona graph for peer alignment, CoPersona effectively mitigates the limitations of weakly observed user attributes. Experimental results across various domains show that CoPersona significantly outperforms existing methods, leading to more robust and accurate personalization of language models.
Sparse user histories can be transformed into rich personalization insights by leveraging collaborative signals from behaviorally similar peers.
Real-world LLM personalization is often constrained by sparse and skewed user histories: most users provide only a handful of interactions, while even frequent users' logs capture an incomplete and biased view of their preferences. As a result, weakly observed user attributes are difficult to infer, leading to brittle personalization when test-time requests shift toward under-supported facets. Motivated by this limitation, we present CoPersona, a graph-based collaborative personalization framework that completes sparse user profiles by borrowing signals from behaviorally similar peers. However, directly transferring signals is difficult because uneven facet coverage introduces bias into interaction histories, obscuring user similarity in the unstructured global space. To address this issue, CoPersona decomposes interaction histories into multiple facet-level representations and explicitly models peer-to-peer, facet-level alignment through a multiplex persona graph. To effectively leverage peer information at inference time, we employ a dual-branch architecture that combines non-parametric peer retrieval with parametric graph reasoning. Experiments across multiple domains and model scales demonstrate consistent improvements over strong baselines, validating CoPersona as an effective approach for robust LLM personalization.
