Reproduce “Improving LLM-powered Recommendations with Personalized Information” on OpenResearch

About this paper

Jiahao Liu, Xueshuo Yan, Dongsheng Li, Guangping Zhang, Hansu Gu, Peng Zhang, Tun Lu, Li Shang and 1 more

CoT-Rec introduces a two-stage pipeline for LLM-powered recommendations that integrates personalized user preferences and item perceptions, derived from Chain-of-Thought processes. This approach enhances both retrieval accuracy and ranking quality while significantly reducing position bias in LLM-as-Ranker models, as demonstrated by improved NDCG and lower MAPB scores across multiple datasets.

Key insights

• Explicitly integrating Chain-of-Thought processes, such as user preference analysis and subjective item perception, enables LLMs to perform more sophisticated and personalized reasoning in recommendation tasks.
• Personalized information, including dynamically updated user preferences and detailed item perceptions, significantly enriches both the retrieval (CRM embedding initialization) and ranking (LLM prompt context) phases of a hybrid recommender system.
• Providing LLMs with structured personalized context via CoT processes can substantially mitigate position bias, a common challenge in LLM-based ranking, by helping the model focus on intrinsic relevance rather than positional cues.

Summary by alphaXiv.