A new paper on arXiv argues that although users express different personalization needs across tasks, these task-specific behaviors can reflect shared underlying preference mechanisms. The authors suggest that existing LLM personalization methods often focus on surface-level behaviors, missing these deeper commonalities.
The paper introduces PersonaFactor, a method that uses preference factors with dynamic selection. Rather than treating each task in isolation, it aims to capture shared preference mechanisms and select the relevant factors for each specific task.
This approach could make personalization more robust and efficient by leveraging cross-task commonalities. The abstract does not include experimental details, so the reported performance and implementation specifics are not yet available from this source alone.