Toward Generative Recommender Systems with Large Language Models: Foundation, Augmentation, and Alignment 3 views

Recommender systems (RS) have become a core component of modern online platforms such as Netflix, LinkedIn, and YouTube. Traditional RSs are generally built on shallow models over interaction data and side information, which increasingly struggle to represent user preferences as user/item semantics grow richer and more complex. Recently, large language models (LLMs) have demonstrated remarkable capabilities in text understanding, reasoning, and generation, motivating a new paradigm of LLM-based generative recommender systems (GRS). Despite this promise, realizing GRS in practice remains non-trivial and raises several fundamental questions: (i) how to faithfully represent users and items within LLMs, (ii) how to incorporate structured and evolving user–item interaction data into LLMs, and (iii) how to systematically leverage user feedback and catalog constraints to align LLM with recommendation objectives. This dissertation addresses these challenges from the following three perspectives: Foundation (Chapter 2): While powerful, pretrained LLMs primarily work in the semantic space of natural language. I develop user/item tokenization methods that integrate textual, behavioral, and relational knowledge into the LLM token space, enabling faithful representations that are compatible with LLM generation. Augmentation (Chapter 3): I propose collaborative augmentation mechanisms to retrieve and incorporate interaction evidence, e.g., collaborative signals and reflective experiences, to ground LLM in real-time user–item behaviors. Alignment (Chapter 4): I introduce catalog-grounded post-training objectives that combine behavior cloning and preference optimization to align LLM generations with user preferences while ensuring feasible, on-catalog recommendations. Collectively, these contributions advance LLM-based GRS through faithful user/item representations, dynamic collaborative knowledge augmentation, and robust alignment under real-world feedback and catalog constraints, bridging the critical gap between LLM capabilities and the practical demands of large-scale recommendation tasks.

PHD (Doctor of Philosophy)

Recommender Systems; Large Language Models; Retrieval Augmented Generation; Supervised Finetuning; Reinforcement Learning

All rights reserved by the author (no additional license for public reuse)

2026-04-21