LUNAR: Benchmarking Personalized Large Language Models on UNiversal User BehAvioR Logs
The LUNAR paper benchmarks how well Large Language Models personalize answers using real user behavior logs across domains like travel and food. Findings show models struggle with evidence retrieval, meaning they cannot effectively find relevant data. The research concludes that improving retrieval methods is key to personalization, while also highlighting the necessary balance between helpfulness and user privacy.