LLM-Based Recommender Systems

Large language models (LLMs) are reshaping recommender systems by bringing deep semantic understanding and text generation into pipelines that traditionally relied on sparse IDs and task-specific models. This survey shows how LLMs improve cold-start accuracy, explanation quality and user engagement, and compares four representative methods (BERT4Rec, P5, TIGER and a headline-generation framework) against a matrix-factorisation baseline. On MovieLens-1M, LLM variants raise Recall@20 by up to 44%, while an online A/B test reports a 7–10% click-through lift from LLM-generated headlines. This paper outlines the trade-off between these gains and a ten-fold rise in inference cost, discusses privacy and carbon-footprint concerns, and argues that hybrid retrieval–generation pipelines and pre-generated content caches will be key to practical deployment. Finally, this paper highlights multimodal dynamic personalisation, e.g. combining adaptive titles and thumbnails — as a promising research frontier.

The paper surveys six paradigms of LLM-based recommendation — sequential transformers (BERT4Rec), the unified text-to-text paradigm (P5), generative retrieval (GPT4Rec), semantic-ID generation (TIGER and LIGER), prompt-based content enrichment (LLM-Rec), and dynamic title personalisation — after first grounding them in the classical collaborative-filtering, content-based and graph-based paradigms they extend.

Why it matters. Most write-ups of LLMs in recommendation stop at “it works better.” This one puts the gain and the bill on the same page: a headline Recall@20 improvement of up to 44% next to a roughly ten-fold increase in inference cost, plus the privacy and carbon consequences that follow from putting a generative model in a hot serving path. That framing is what a team actually needs to decide whether to adopt any of these methods, and it leads to the concrete architectural recommendation the paper defends — hybrid retrieval–generation with pre-generated content caches, rather than generation at request time.