How Large Language Models are Transforming Recommendation Systems: Insights from Huawei

Overview

Huawei Noah's Ark Lab presented a comprehensive study on applying large language models (LLMs) to recommendation systems, discussing data, model, and process perspectives, and introducing two key projects—KAR and Uni-CTR—that aim to enhance recommendation performance and user experience.

Background and Problem

Traditional recommendation systems rely on closed‑loop data such as user interactions and item attributes, limiting their ability to incorporate real‑world knowledge. This results in suboptimal performance, especially for cold‑start items or sparse data scenarios.

LLM4Rec Exploration and Application

KAR: Knowledge‑Assisted Recommendation

KAR (Knowledge‑Assisted Recommendation) uses LLMs to inject open‑domain knowledge into recommendation pipelines. By generating user‑preference and item‑fact reasoning through templated prompts, the approach produces dense knowledge vectors that are fused with traditional embeddings via a multi‑expert network, improving robustness and reducing hallucination.

Uni-CTR: Multi‑Scenario Recommendation Base

Uni-CTR tackles cross‑domain recommendation by converting structured tabular data into natural‑language prompts for LLMs, extracting both scene‑specific and shared representations through leader and backbone networks. The method supports zero‑shot cold‑start scenarios and demonstrates significant gains across fashion, music, and gift‑card datasets.

Experimental Results

Both KAR and Uni-CTR show notable AUC improvements (≈1%) over baseline models. KAR achieves comparable inference latency to base models, making it viable for industrial deployment. Uni-CTR delivers balanced performance across multiple scenes, avoiding the trade‑offs typical of single‑scene optimizations.

Challenges and Outlook

Key challenges include joint modeling of collaborative and semantic signals, efficient handling of long textual inputs and ID encodings, and meeting real‑time inference constraints in production environments.

Future work will focus on enriching recommendation data with world knowledge, shifting from discriminative to generative models, and unifying multi‑stage pipelines into end‑to‑end LLM‑driven recommendation systems.