How Huawei Noah’s KAR Project Leverages LLMs to Advance Recommendation Systems

The paper provides a systematic review of recommendation system technology, tracing its progression from early deep‑learning approaches to the current era of large language models (LLM) and AI agents. It identifies three persistent challenges in traditional systems: high noise in implicit feedback, insufficient semantic comprehension, and difficulty extracting user intent.

To address these issues, the authors examine two recommendation paradigms—list‑based and conversational—and propose integrating LLMs as feature enhancers within the recommendation pipeline. Using Huawei Noah’s Ark (KAR) project as a concrete case, the article explains how factorized prompting and a multi‑expert knowledge adapter map semantic knowledge into the recommendation embedding space efficiently.

The design of the multi‑expert adapter is highlighted, showing how it balances the dimensionality of textual features with the real‑time constraints of recommendation models. The authors also discuss prompting engineering and fine‑tuning strategies for dialog‑based recommendation, as well as a multi‑capability AI‑Agent architecture that coordinates various abilities across the system.

Experimental results are presented, including a quantitative AUC improvement of 1.5% and supporting online A/B testing data, demonstrating the practical impact of the proposed techniques. The article concludes with a forward‑looking discussion on cross‑platform recommendation ecosystems and future research directions.