Understanding Generative Recommendation with Semantic IDs from a Model-scaling View
A new research paper explores the limitations of generative recommendation systems that use semantic IDs, finding their performance saturates as models scale up. The study proposes that directly using large language models (LLMs) as recommenders offers better scaling properties and can achieve up to 20% performance improvement. This research suggests LLM-as-RS is a more promising direction for future generative recommendation foundation models. AI
IMPACT Suggests LLM-based recommendation systems scale better than current semantic ID approaches, potentially improving user experience.