保留物品语义：重新思考LLM生成式推荐中的Token初始化
原标题：Preserving Item Semantics for Free: Rethinking Token Initialization in LLM-Based Generative Recommendation
AI 摘要
该研究针对基于LLM的生成式推荐系统中语义ID（SID）的初始化问题，发现标准随机初始化会导致嵌入围绕物品流行度而非语义组织，且持续预训练（CPT）无法可靠恢复语义几何。作者提出一种零参数干预方法，直接从语义嵌入空间的质心初始化SID token嵌入，无需额外训练或推理开销，在纯SFT下Recall@5提升最高16%，冷物品Recall@5提升最高60%，并减少SFT和CPT训练步数。该工作表明保留SID几何结构可为LLM生成式推荐提供简单有效的语义先验。
正文节选
Computer Science > Information Retrieval Title:Preserving Item Semantics for Free: Rethinking Token Initialization in LLM-Based Generative Recommendation View PDF HTML (experimental) Abstract:Recent advances in generative recommendation (GR) leverage large language models (LLMs) as recommender backbones, enabling LLMs to directly generate recommendations conditioned on item-interaction histories. In these systems, items are often represented through semantic IDs (SIDs) added to the L