Understanding Generative Recommendation with Semantic IDs from a Model-scaling View

Researchers demonstrate that semantic ID-based generative recommendation systems hit significant scaling bottlenecks, while large language models used directly as recommenders show superior scaling properties and up to 20% performance improvements. This challenges current approaches in generative recommendation and suggests LLM-based systems represent a more promising path forward for recommendation foundation models.

The research addresses a critical inefficiency in modern recommender systems that attempt to blend rich semantic information with collaborative filtering. Semantic ID-based approaches—which quantize embeddings from language and vision models into discrete tokens—have become popular but exhibit performance saturation when scaled, revealing fundamental architectural limitations. The study reveals that as model components enlarge, improvements plateau quickly, suggesting the discrete tokenization process loses semantic information necessary for effective recommendations.

This finding emerges against the backdrop of rapid advances in generative AI, where scaling laws have consistently driven improvements across language and vision domains. The research team's pivot toward LLM-as-RS models—using large language models directly without semantic ID intermediaries—demonstrates fundamentally different scaling behavior. Testing across model sizes from 44M to 14B parameters shows LLMs progressively improve at capturing user-item interaction patterns, contradicting assumptions that these models lack collaborative filtering capabilities.

The implications extend beyond academic interest. Companies investing in semantic ID-based recommendation infrastructure may face diminishing returns on scaling investments, while those exploring direct LLM deployment could achieve competitive advantages. The research suggests that foundation models for recommendation systems may require rethinking current architectural assumptions about tokenization and information encoding.

Looking forward, the industry should monitor whether these findings hold across diverse datasets and recommendation contexts, particularly in e-commerce and content discovery platforms where both semantic richness and personalization matter. The path toward more capable recommendation foundations likely involves deeper integration with LLM architectures rather than intermediate tokenization schemes.

• →Semantic ID-based generative recommendation systems experience performance saturation when scaled, indicating fundamental capacity limitations in discrete tokenization approaches.
• →LLM-as-RS paradigm achieves up to 20% improvement over semantic ID methods through scaling and demonstrates superior model scaling properties.
• →Large language models effectively capture collaborative filtering signals despite prior assumptions they would struggle with interaction modeling.
• →Research spans 44M to 14B parameter models, providing robust evidence across multiple scales of the architectural advantages of direct LLM-based recommendations.
• →Foundation models for recommendation systems may require architectural rethinking away from intermediate tokenization toward direct LLM integration.