LLM recommender rewards cause exposure collapse, new study finds

Researchers have identified a key factor contributing to exposure collapse in large language model (LLM) recommenders: the item-mean component of collaborative-filtering (CF) scores. This component, representing an item's average score across all users, can lead LLMs to concentrate on a limited set of items, even if the underlying CF model offers diverse recommendations. To address this, the study proposes two reward correction methods: per-item centering and per-item standardization. Experiments demonstrated that these corrections significantly improve catalog coverage and exposure evenness, increasing the number of distinct items ranked first by up to 9.8 times while maintaining ranking accuracy. AI

IMPACT Improves the diversity and coverage of recommendations generated by LLM-based systems.

RANK_REASON Academic paper detailing a novel method for improving LLM recommender systems. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.IR (Information Retrieval) →

• arXiv
• collaborative filtering
• CORE Recommender
• DagsHub
• Hugging Face
• large language model

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →