How Kuaishou Boosted Ad Performance with Multimodal Large Models

Overview

In the past two years Kuaishou has explored the use of large‑model technology in its advertising scenario, addressing data and model design motivations and challenges.

Challenges in Content‑Domain Ad Estimation

Kuaishou’s platform combines content and e‑commerce, offering various media (images, short videos, live streams, product pages). User behavior is scattered across these media, making single‑scene data sparse. The existing ID‑centric recommendation system (video ID, item ID, live ID) hinders cross‑domain behavior integration, and differing content lifecycles further complicate modeling.

Leveraging Global Behavior with COPE

To unify product content representation, Kuaishou proposes the COPE framework, which creates a unified embedding for items across modalities, enabling more effective use of whole‑platform user actions.

Injecting External Knowledge with LEARN

The LEARN framework transfers knowledge from large language models (LLMs) to the ad system, using LLMs’ world knowledge and reasoning abilities to break the “information cocoon” and improve ad relevance.

By integrating multimodal large models and LLM‑based knowledge transfer, Kuaishou achieves measurable business gains in ad performance.