REGER: Modeling inter-item Relations with Semantic IDs for Generative Recommendation
生成式推荐模型如何引入各种关系, 尤其给出判别式模型 -> 树模型的概率转换公式
Blogs
Neither Black nor White: Balancing Semantic and Collaborative Signals with Graph-Informed Semantic IDs (GrIS)
GrIS, 将 SID 构造拆成协同图构建与层级划分, 通过 RecDMoN 和 RQ-GAE 融合内容与行为信号
HASTE: Hierarchical and Attribute-aware Supervision for Generative Recommendation
HASTE, 结合层级加权、置信度门控与属性回归监督, 改善 SID 生成和长尾推荐
S²GR: Stepwise Semantic-Guided Reasoning in Latent Space for Generative Recommendation
S²GR, 在每一位 SID 生成前插入潜在思考向量, 以码本聚类语义提供逐步监督
Tokenize Once, Recommend Anywhere: Unified Item Tokenization for Multi-domain LLM-based Recommendation
UniTok, 利用量化专家与域间信息校准构建可跨域复用的 item tokenizer
Universal Item Tokenization for Transferable Generative Recommendation
UTGRec, 联合学习多模态内容编码与离散化, 通过共享码本迁移生成式推荐
Guiding the coarse levels of semantic IDs makes the fine levels learnable
Guided SID, 将任务相关属性直接指定为粗层编码, 让残差量化在属性划分后学习细节
What Makes a Good Semantic ID for Generative Recommendation? A Reproducibility Study
SID-Repro, 从推荐效果、码本利用率、扩展行为和局部语义保留评估 Semantic ID
An Efficient Private GPT Never Autoregressively Decodes
POST, 用客户端公开模型起草与安全批量验证摊薄私有 GPT 的密码学推理成本
Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads
用多个轻量预测头并行提出未来 token, 通过树状验证减少解码轮数