How Kuaishou Boosted Ad Performance with Multimodal LLMs: COPE & LEARN Frameworks

Introduction

This article reviews Kuaishou’s exploratory work over the past two years on applying large‑model technologies to advertising scenarios, describing the data and model design motivations.

Key Challenges and Opportunities

Kuaishou’s platform combines content and e‑commerce, with user behavior scattered across multiple media (short videos, live streams, product pages, etc.) and IDs (video ID, item ID, live ID). This leads to sparse ad data and cross‑domain modeling difficulties, as well as information‑filter bubbles.

Full‑Domain Behavior Utilization – COPE Framework

To unify product representation across scenes, Kuaishou introduced the COPE (Content‑Unified Representation) framework. It moves beyond simple item IDs by aggregating semantically related items into SPU IDs and compressing multimodal content from videos, live streams, and detail pages into a robust, compact feature vector, improving item‑side features and reducing reliance on sparse behavior.

External Knowledge Transfer – LEARN Framework

Leveraging large language models (LLMs), the LEARN framework injects world knowledge and reasoning ability into ad recommendation models, helping break the information‑cocoon effect and enhancing user interest modeling.

Both frameworks have delivered tangible business improvements on Kuaishou’s advertising system.