What WWW'24 Papers Reveal About LLMs in Search & Recommendation

This article presents a concise survey of six papers from WWW 2024 that explore the use of large language models (LLMs) in industrial search and recommendation scenarios.

Alibaba: Large Language Model based Long‑tail Query Rewriting in Taobao Search

The authors introduce BEQUE, a three‑stage framework consisting of multi‑instruction supervised fine‑tuning (SFT), offline feedback, and target alignment. They construct a rewrite dataset via rejection sampling and auxiliary tasks, fine‑tune an LLM, generate multiple candidates with beam search, and feed them to Taobao’s offline system for partial ranking. Contrastive learning highlights differences among rewrites, and the model is aligned with online business goals. Offline experiments show effective semantic gap reduction, while online A/B tests report significant gains in GMV, transaction volume, and visitor count. BEQUE has been deployed on Taobao since October 2023.

Microsoft: Knowledge‑Augmented Large Language Models for Personalized Contextual Query Suggestion

The paper proposes a method that enriches LLM prompts with user‑centric knowledge stored in a lightweight entity‑based repository built from search and browsing logs. This repository aggregates user interests onto a public knowledge graph, preserving privacy and scalability. A soft‑plus‑hard prompting strategy combines user/item tokens with vocabulary tokens, enabling the LLM to generate context‑aware query suggestions. Human‑evaluated experiments demonstrate superior relevance, personalization, and usefulness compared to baseline LLM approaches.

Alibaba: Modeling User Viewing Flow using Large Language Models for Article Recommendation

The SINGLE method models both static and instant user viewing flows. Static modeling captures a user’s general interests, while instant modeling focuses on recent clicks to reflect immediate intent. LLMs extract persistent preferences from previously clicked articles and adaptively attend to candidate article representations. Offline experiments on Alibaba’s ATA platform and online A/B testing show a 2.4% improvement over prior baselines.

LinkedIn: Collaborative Large Language Model for Recommender Systems (CLLM4Rec)

CLLM4Rec tightly integrates the LLM paradigm with traditional recommender‑system (RS) paradigms. User/item IDs extend the LLM vocabulary, and a novel soft+hard prompting scheme mixes collaborative and content tokens. A bidirectional regularization encourages the model to capture recommendation‑relevant signals from noisy textual data. Finally, a polynomial‑likelihood recommendation head predicts retained items based on soft+hard prompts, enabling efficient multi‑item generation without hallucination.

Baidu: Representation Learning with Large Language Models for Recommendation (RLMRec)

RLMRec is a model‑agnostic framework that enhances existing ID‑based recommenders with LLM‑derived textual signals. It incorporates auxiliary text, uses LLMs to model user/item features, and aligns the LLM semantic space with collaborative signals via cross‑view alignment. Maximizing mutual information improves representation quality, and experiments demonstrate robustness and efficiency on noisy data compared with state‑of‑the‑art baselines.

Huawei: Retrieval‑enhanced Large Language Models for Lifelong Sequential Behavior Comprehension (ReLLa)

ReLLa addresses zero‑shot and few‑shot recommendation challenges by augmenting LLMs with a semantic user‑behavior retrieval module (SUBR). For zero‑shot tasks, SUBR improves test sample quality; for few‑shot tasks, Retrieval‑enhanced Instruction Tuning (ReiT) creates a hybrid training set mixing original and retrieved samples. Experiments on three public datasets show ReLLa outperforms strong baselines, and the few‑shot variant achieves higher CTR than full‑data models such as DCNv2, DIN, and SIM while using less than 10% of the training data.

Other WWW'24 LLM‑related Works

ClickPrompt: CTR Models are Strong Prompt Generators for Adapting Language Models to CTR Prediction

Search‑in‑the‑Chain: Interactively Enhancing Large Language Models with Search for Knowledge‑intensive Tasks

Enhancing sequential recommendation via LLM‑based semantic embedding learning

NoteLLM: A Retrievable Large Language Model for Note Recommendation

Harnessing Large Language Models for Text‑Rich Sequential Recommendation

Back to the Future: Towards Explainable Temporal Reasoning with Large Language Models

LinkNER: Linking Local Named Entity Recognition Models to Large Language Models using Uncertainty

Labor Space: A Unifying Representation of the Labor Market via Large Language Models

Explainable Fake News Detection With Large Language Model via Defense Among Competing Wisdom

Metacognitive Retrieval‑Augmented Large Language Models

Cognitive Personalized Search Integrating Large Language Models with an Efficient Memory Mechanism

PMG: Personalized Multimodal Response Generation with Large Language Models

Mechanism Design for Large Language Models

Learning to Generate Explainable Stock Predictions using Self‑Reflective Large Language Models

Unifying Local and Global Knowledge: Empowering Large Language Models as Political Experts with Knowledge Graphs

KGQuiz: Evaluating the Generalization of Encoded Knowledge in Large Language Models

Ask Me in English Instead: Cross‑Lingual Evaluation of Large Language Models for Healthcare Queries

Harnessing Multi‑role Capabilities of Large Language Models for Open‑domain Question Answering

GraphTranslator: Aligning Graph Model to Large Language Model for Open‑ended Tasks

Query in Your Tongue: Reinforce Large Language Models with Retrievers for Cross‑lingual Search Generative Experience

Conclusion

LLM‑enhanced search and recommendation remain concentrated in text‑driven scenarios such as news, article, and dialogue recommendation, as well as in collaborative‑filtering and graph‑based representation learning. Applications to query rewriting and contextual suggestion are also prominent. However, no breakthrough work yet reshapes classic paradigms or achieves massive real‑world deployment, suggesting ample opportunity for novel contributions in 2024 and beyond.