Integrating Large Language Models with Recommendation Systems: Paradigms, Methods, and Experiments

This article reviews and discusses the integration paradigm of LLM + Recommendation, attempting to transfer the emergent abilities of large language models (LLMs) to recommendation systems. By leveraging the general knowledge of LLMs, the goal is to assist recommendation, improve effectiveness, and enhance user experience.

Background : E‑commerce recommendation systems (RecSys) personalize item suggestions based on user behavior and preferences. Traditional RecSys rely on ID‑based features, lacking semantic and external knowledge, which leads to cold‑start, filter‑bubble, and diversity issues. LLMs, trained on massive corpora, exhibit emergent and generalization abilities, storing world knowledge and providing strong language understanding.

Two modeling paradigms :

1. LLM + Recommendation – LLM acts as a feature extractor. Raw item/user information (titles, attributes, click sequences) is formatted as prompts and fed to the LLM, which outputs embeddings or semantic summaries. These features are then incorporated into downstream recommendation modules.

2. LLM as Recommendation – The pre‑trained LLM directly replaces one or more stages of the RecSys pipeline (recall, coarse‑ranking, fine‑ranking, re‑ranking). The model receives prompts containing user context and task instructions and outputs final recommendation results.

LLM + Recommendation implementations :

• LLM Embedding : Use the LLM as an encoder to obtain dense embeddings for item titles/attributes, e.g., U‑BERT and UniSRec.

• LLM Summary : Summarize raw item/user text into concise semantic statements for downstream use, e.g., the GENRE framework for news recommendation.

• LLM as Ranker : Formulate ranking as a conditional generation task; prompts combine user history, candidate items, and ranking instructions. Zero‑shot LLMs can score candidates, though position and popularity bias remain.

• LLM as RecSys (Chat‑Rec) : Multi‑turn dialogue between user and LLM narrows candidate sets and provides explanations.

Algorithmic proposals :

• Use LLM knowledge to construct category‑matching systems, enriching recall and ranking with cross‑category recommendations.

• Apply LLM for text denoising and rewriting of noisy product titles, then embed the cleaned text (e.g., using CoROM) and fuse with ID‑based features in a dual‑tower DSSM.

• Extend ranking features with LLM‑derived category embeddings, improving CTR and diversity.

Experimental results :

• Adding LLM‑derived semantic vectors to the item tower improves AUC by +0.064 pt compared to baseline.

• Category‑matching via LLM boosts uCTR (+0.83 %), average IPV (+2.58 %), and leaf‑category clicks (+2.06 %).

• Regularizing product titles with LLM reduces perplexity, indicating more coherent semantics.

Future directions :

Multimodal recommendation: incorporate image, video, and audio cues via multimodal LLMs.

LLM inference acceleration: distillation, pruning, quantization to meet millisecond‑level latency.

LLM as re‑ranking: leverage LLM knowledge to select final items from top‑k lists.

Code example (prompt for product description generation) :

你现在是一个买家。给定商品的描述词【A】以及各种属性【B】，请根据关键词和关键属性描述出商品是什么。要求是只需要回答是什么，不要补充其他内容，尽量从A和B中选出词语进行描述，字数不超过40，回答模版为:这个商品是...。比如当A=['giyo', '公路', '山地车', '专用', '自行车', '单车', '专业', '骑行', '手套', '半指', '夏季', '男', '硅胶', '减震', '女']，B=['尺码': 'XXL', '类目': '自行车手套', '适用对象': '通用', '颜色分类': '弧光半指-黄色-双面透气+GEL硅胶+劲厚掌垫', '上市时间': '2016年夏季', '货号': '1183', '品牌': 'GIYO/集优', '款式': '半指手套']，输出：这个商品是GIYO牌的自行车半指手套。现在A=...,B=...

The authors conclude that LLMs will evolve from auxiliary tools to core components of recommendation systems, driving richer semantic understanding, multimodal integration, and more personalized, explainable recommendations.