SpecInfer: Accelerating Large Language Model Serving with Tree-based Speculative Inference and Verification
用 token tree 并行验证多条推测路径以加速 LLM 服务
Blogs
SpecTr: Fast Speculative Decoding via Optimal Transport
从最优传输理解投机解码及其多候选扩展
Fast Inference from Transformers via Speculative Decoding
用小模型猜测, 大模型并行验证, 且保持精确采样
SEQUOIA: Scalable and Robust Speculative Decoding
用动态规划优化 token tree, 以无放回采样提高验证鲁棒性
Rethinking Generative Recommender Tokenizer: Recsys-Native Encoding and Semantic Quantization Beyond LLMs
从信息论角度剖析 SID 的设计理念
A Rank-One Popularity Component in Dot-Product Recommender Scores
推荐 Softmax 目标中的 Rank-One Popularity Component
R3-VAE: Reference Vector-Guided Rating Residual Quantization VAE for Generative Recommendation
更稳定的生成式推荐 Tokenizer: 正交残差分量 & Cosine & 聚类正则
HiD-VAE: Interpretable Generative Recommendation via Hierarchical and Disentangled Semantic IDs
用层次化解耦 VAE 生成可解释的语义 ID
Markovian Pre-Trained Transformer for Next-Item Recommendation
用合成马尔科夫链预训练推荐模型
BEAR: Beam-Search-Aware Optimization for LLM-based Recommendation
Beam Search 的训推一致约束学习