Towards Effective Model Editing for LLM Personalization

ICLR 2026 Conference Submission14658 Authors

Keywords: Personalization, Model Editing, LLMs

Abstract: Customer-facing LLMs need personalization to reflect individual preferences and needs. However, existing personalization methods are often computationally expensive, data-intensive, prone to catastrophic forgetting, and degrade in multi-turn conversations and on implicit questions. To address these challenges, we conceptualize personalization as model editing and present Personalization Editing, a framework that applies localized edits guided by clustered preference representations, enforcing desired behavior where preferences apply while preserving other capabilities. Existing personalization datasets often use synthetic personas in role-playing dialogues, leading to indirect evaluation that does not reflect real-world user queries. We introduce UPQA, a short-answer QA dataset based on in-situ user queries, with varying levels of difficulty. Unlike prior benchmarks, UPQA directly tests whether models can recall and apply specific user preferences, enabling more accurate and efficient evaluation. Across settings, Personalization Editing improves editing accuracy, and is more computationally efficient than fine tuning, while outperforming prompting and retrieval based baselines in multi-turn conversations and on implicit preference questions.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 14658