LEADRE: Knowledge‑Enhanced LLMs Supercharge Display Ad Recommendations

Paper Overview

The paper “LEADRE: Multi‑Faceted Knowledge Enhanced LLM Empowered Display Advertisement Recommender System” was selected for the VLDB 2025 Industry Track. It is a joint effort by Tencent Advertising’s commercial recommendation team and Renmin University of China, presenting a generative recall solution based on the “混元” large language model.

Challenges in Display Advertising

Three key challenges are identified:

Accurately modeling user interest in weak‑intent, sparse‑behavior scenarios.

Bridging the knowledge gap between general‑purpose LLMs and the domain‑specific advertising knowledge.

Deploying LLM‑driven generative recall at billions of daily requests with low latency.

LEADRE Architecture

LEADRE consists of three core components:

Intent‑aware Prompt Engineering.

Advertising‑Domain Knowledge Alignment.

Latency‑Aware Model Deployment.

It is built on a 1‑billion‑parameter “混元” model and has been deployed in Tencent’s display ad system, serving billions of users and handling hundreds of billions of requests per day.

Intent‑Aware Prompt Engineering

The system constructs <Prompt, Response> pairs from user behavior sequences and ad descriptions. Prompts combine long‑term interest (user profile, historical actions) and short‑term interest (recent interactions) and include sections such as Task Description, User Profile, Interest Summary, Advertising Behavior Sequence, and Content Behavior Sequence. Multiple prompt templates, user‑profile re‑ordering, and positive‑interaction reuse are employed to enrich training data.

Advertising‑Domain Knowledge Alignment

Two auxiliary tasks are introduced:

Explicit alignment: Predict S‑IDs from detailed ad descriptions to teach the LLM the mapping between text and internal identifiers.

Implicit alignment: Replace S‑IDs with ad descriptions in the next‑ad generation task, forcing the model to reason over textual semantics.

Direct Preference Optimization (DPO) is applied to bias the model toward higher‑value ads based on ECPM.

Latency‑Aware Deployment

The system separates inference into a Latency‑Sensitive Service for immediate requests and a Latency‑Tolerant Service that performs near‑line LLM inference using a Trie‑Tree and TensorRT acceleration. Training updates the LLM daily from user‑feature databases.

Experimental Results

In A/B tests covering 20 % of traffic on WeChat Channels and Moments, LEADRE increased GMV by 1.57 % and 1.17 % respectively. Adding retrieved ads as features further boosted GMV by 1.43 % on Channels.

Conclusion and Future Work

The study demonstrates that a knowledge‑enhanced LLM can effectively improve recall quality and business metrics in large‑scale display advertising. Future work will extend generative techniques from recall to the ranking stage.