How Kuaishou Boosted Ad Performance with Multimodal Large Models: COPE & LEARN

Introduction

In the past two years Kuaishou has investigated how large‑model and multimodal techniques can improve advertising estimation within its content‑driven platform. The article outlines the challenges of ad prediction, the motivation for leveraging whole‑platform user behavior and external knowledge, and introduces two key frameworks: the COPE unified product representation and the LEARN LLM‑based knowledge transfer.

Challenges and Opportunities

Kuaishou’s ecosystem mixes short videos, live streams, product detail pages and landing pages, generating user actions that are sparse and fragmented across many media and ID spaces (Video ID, Item ID, Live ID, etc.). Traditional recommendation pipelines rely on ID‑centric embeddings, which lack semantic meaning and hinder cross‑domain interest modeling. Moreover, recommendation loops create “information cocoons” that limit the introduction of new knowledge.

Global Behavior Utilization – COPE

To break the ID barrier, Kuaishou first built a SPU (Standard Product Unit) ID that aggregates items with the same attributes, providing a more stable identifier. However, SPU IDs still depend heavily on interaction data. The COPE framework compresses multimodal product content (images, text, video) into a robust, compact embedding that can be shared across short‑video, live‑stream and e‑commerce scenarios, enriching item features and reducing reliance on sparse behavior signals.

External Knowledge Transfer – LEARN

Large language models (LLMs) are employed to inject world knowledge and reasoning ability into the ad system. By fine‑tuning or prompting LLMs, Kuaishou transfers external knowledge to break the information cocoon, improving ad relevance and business metrics.

Results

Integrating COPE and LEARN has yielded measurable business improvements in ad click‑through rates and conversion, demonstrating the practical value of multimodal large‑model techniques in a real‑world advertising platform.