How Kuaishou Boosted Ad Performance with Multimodal LLMs and the COPE Framework

Introduction

This article outlines Kuaishou’s exploratory work over the past two years that applies large‑model technology to advertising scenarios, describing the data and model design motivations behind the efforts.

1. Content‑domain ad estimation challenges and large‑model opportunities

Kuaishou’s platform combines content and e‑commerce, offering various media forms such as images, short videos, live streams, product detail pages, and landing pages. User behavior is scattered across these heterogeneous contexts, making single‑scene data sparse. Existing recommendation systems rely on ID‑centric pipelines (Video ID, Item ID, Live ID, etc.), which hinder cross‑domain data sharing and complicate interest modeling across different content lifecycles.

Advertising data is especially sparse, but integrating full‑domain behavior—both natural‑flow and search interactions—can improve user interest modeling. Additionally, recommendation models often fall into “information cocoons” by repeatedly training on historical user actions; injecting external knowledge via large language models (LLMs) can help break this loop.

2. Full‑domain behavior utilization: the COPE unified product‑content representation framework

To leverage full‑domain behavior, Kuaishou first built a product‑ID system (SPU ID) that aggregates semantically similar items across scenes, providing a more stable identifier than raw Item IDs. However, SPU IDs still lack semantic richness. The COPE framework compresses multimodal product content from live streams, detail pages, short videos, etc., into a robust, compact feature representation. This unified representation enhances item‑side features, expands user behavior sequences, and reduces dependence on scene‑specific interaction data.

3. External knowledge utilization: the LEARN LLM knowledge‑transfer framework

To address the information‑cocoon problem, Kuaishou employs the open‑pretrained capabilities of large language models. By transferring world knowledge and strong reasoning abilities from LLMs, the LEARN framework injects external knowledge into the advertising recommendation pipeline, helping break the feedback loop and improving model generalization.

Both COPE and LEARN have delivered measurable business gains in Kuaishou’s ad system.

Article excerpted from the e‑book “A Plain‑spoken Large‑Model Handbook”.