arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2511.11255cs.IR
收藏

Align$^3$GR: 基于大语言模型的生成推荐系统统一多级对齐

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

Kuaishou Technology(快手科技)

Wencai Ye, Mingjie Sun, Shuhang Chen, Wenjin Wu, Peng Jiang

AI总结:

Align$^3$GR通过统一多级对齐方法,提升大语言模型在生成推荐系统中的性能,实验显示在Recall@10和NDCG@10指标上均取得显著提升。

摘要:

大型语言模型(LLMs)在利用结构化世界知识和多步推理能力方面表现出显著优势。然而,将LLMs转换为现实世界推荐系统时,由于语义和行为不匹配的根本性挑战而产生困难。为此,我们提出了Align$^3$GR,一种新的框架,统一了token级、行为建模级和偏好级对齐。我们的方法引入了双token化融合用户-物品语义和协同信号。增强的行为建模通过双向语义对齐。渐进式DPO策略结合自我扮演(SP-DPO)和现实反馈(RF-DPO)以实现动态偏好适应。实验显示,Align$^3$GR在公开数据集上比SOTA基线在Recall@10上提高了17.8%,在NDCG@10上提高了20.2%,并在在线A/B测试和工业级大规模推荐平台的全规模部署中取得了显著收益。

英文摘要:

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.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑