Align$^3$GR: Unified Multi-Level Alignment for LLM-based Generative Recommendation

TL;DR

Align$^3$GR introduces a comprehensive multi-level alignment framework for LLM-based recommendation systems, significantly improving their accuracy and adaptability by unifying semantic, behavioral, and preference alignments.

Contribution

This paper presents a novel unified alignment framework that integrates token, behavior, and preference levels to enhance LLM-based recommender systems.

Findings

Outperforms SOTA by +17.8% in Recall@10

Achieves +20.2% in NDCG@10

Demonstrates significant online and industrial deployment gains

Abstract

Large Language Models (LLMs) demonstrate significant advantages in leveraging structured world knowledge and multi-step reasoning capabilities. However, fundamental challenges arise when transforming LLMs into real-world recommender systems due to semantic and behavioral misalignment. To bridge this gap, we propose Align$^3$GR, a novel framework that unifies token-level, behavior modeling-level, and preference-level alignment. Our approach introduces: Dual tokenization fusing user-item semantic and collaborative signals. Enhanced behavior modeling with bidirectional semantic alignment. Progressive DPO strategy combining self-play (SP-DPO) and real-world feedback (RF-DPO) for dynamic preference adaptation. Experiments show Align$^3$GR outperforms the SOTA baseline by +17.8% in Recall@10 and +20.2% in NDCG@10 on the public dataset, with significant gains in online A/B tests and full-scale deployment on an industrial large-scale recommendation platform.