arxiv:2601.11000

When Personalization Misleads: Understanding and Mitigating Hallucinations in Personalized LLMs

Published on Jan 16

· Submitted by

SunZX on Jan 19

Renmin University of China

Upvote

Authors:

Zhongxiang Sun ,

Abstract

Personalized large language models can generate false information aligned with user history instead of factual truth, but a new method called FPPS helps maintain both factual accuracy and personalized responses while preserving existing personalization effects.

AI-generated summary

Personalized large language models (LLMs) adapt model behavior to individual users to enhance user satisfaction, yet personalization can inadvertently distort factual reasoning. We show that when personalized LLMs face factual queries, there exists a phenomenon where the model generates answers aligned with a user's prior history rather than the objective truth, resulting in personalization-induced hallucinations that degrade factual reliability and may propagate incorrect beliefs, due to representational entanglement between personalization and factual representations. To address this issue, we propose Factuality-Preserving Personalized Steering (FPPS), a lightweight inference-time approach that mitigates personalization-induced factual distortions while preserving personalized behavior. We further introduce PFQABench, the first benchmark designed to jointly evaluate factual and personalized question answering under personalization. Experiments across multiple LLM backbones and personalization methods show that FPPS substantially improves factual accuracy while maintaining personalized performance.

View arXiv page View PDF Add to collection

Community

Yi-Zhan

about 5 hours ago

💡 Overview
Personalization is increasingly adopted in modern LLM systems, but we find it can systematically distort factual reasoning. We identify personalization-induced hallucinations, where models generate answers aligned with user history rather than objective truth. To mitigate this, we also propose Factuality-Preserving Personalized Steering (FPPS), a lightweight inference-time approach that detects harmful personalization and adaptively steers internal representations to recover factual correctness while keeping useful personalization.

🔥Key Insights

Problem discovery: We provide the first systematic study of personalization-induced hallucinations and show risks to factual reliability, downstream knowledge acquisition, and long-term user trust.
Mitigation method: We propose FPPS, a lightweight inference-time framework that selectively restores factuality under personalization.
Evaluation dataset: We develop PFQABench to jointly evaluate factual QA and personalized QA under aligned user sessions, enabling controlled assessment of factuality failures and mitigation.
Results: Extensive experiments across multiple LLM backbones and personalization methods show FPPS substantially improves factual accuracy without sacrificing personalization performance.