Large Language Models for Recommendation Systems: Current Progress, Challenges, and Future Directions

This tutorial provides an overview of how large language models (LLMs) are being applied to recommendation systems, covering background concepts, recent progress, open challenges, and future research directions.

Background of RecSys – A simplified recommendation workflow is illustrated, showing how multi‑stage filtering and user feedback form a feedback loop for model training.

Benefits of LMs – LLMs bring strong interaction, generalization, and generation capabilities that can enhance user experience, cross‑domain recommendation, and content creation.

Progress of LLM4Rec – The field is organized into three dimensions: metrics (accuracy, trustworthiness, fairness, privacy, safety, OOD), information modalities (text, image, video), and technical approaches (in‑context learning, tuning, agents).

In‑context Learning – Describes prompt‑based methods for point‑wise, pair‑wise, and list‑wise ranking, as well as data augmentation via LLM‑generated user profiles.

Tuning LLM4Rec

TALLRec – Partial‑parameter fine‑tuning (e.g., LoRA) on a 7B LLaMA model using few samples to quickly adapt to new recommendation tasks.

InstructRec – Full‑parameter tuning with diverse instruction data for tasks such as product search, personalized search, and sequential recommendation.

BIGRec – Generative approach that directly predicts the next item title, highlighting challenges of tokenizing items and grounding generated titles.

TransRec – Constrained generation using multi‑facet item identifiers and FM‑index to ensure valid token sequences.

LC‑Rec – Uses codebook‑based item indexing via auto‑encoders to capture multimodal item information.

Agent for Recommendation – Discusses agents as user simulators for interactive evaluation and agents that directly perform recommendation by leveraging LLM planning, tool use, and multi‑turn interaction.

Open Problems – Includes modeling challenges for relational recommendation data, cost of training and inference, evaluation metrics for long‑term and interactive recommendation, and data issues such as the need for new, semantically rich benchmark datasets.

Future Directions & Conclusions – Highlights generative recommendation paradigms, Rec4Agentverse for recommending AI agents, and the importance of AI‑generated content for enriching recommendation ecosystems.

Q&A – Addresses practical questions about LLM usage in industry, data scale, and how LLMs can motivate higher‑quality user‑generated content.