第 171 期 · 2026 年 10 月 24 日 周六
为搜索与推荐工程师挑选的每日 GenAI 论文 · 评分 · 深度解读Generative · 生成式推荐 — 7 篇入选
头条:UniRec: A Unified Expressive-Aligned Generative Recommendation Framework for E-commerce — Shopee 的作者统计了标准 RQ-KMeans 生成的三层语义 ID(SID):最常曝光的前 10% 完整编码拿走了 89.62% 的曝光,生成式推荐模型的概率因此集中在少数热门编码上。UniRec 让模型先生成商品的二级、三级类目 token,再生成三层 SID,并在构建 SID 时给每个编码加一个按曝光加权的容量上限。先生成类目,后面每一层 SID 只需要在这个类目里挑。离线 HR@50 比 OneRec-V2 相对高 22.6%;线上对比原多阶段链路,订单 +4.76%,GMV +5.60%。
- UniRec: A Unified Expressive-Aligned Generative Recommendation Framework for E-commerce(深度解读)
- OneLatent: Latent Reasoning for Efficient Foundation Recommendation Models(深度解读)
- GenPage: Towards End-to-End Generative Homepage Construction at Netflix(深度解读)
- Coarse-to-Fine Long-term Interest Modeling for Generative Recommendation(深度解读)
- CSG4Rec: Collaborative-Similarity Guided Semantic IDs Tokenization for Generative Sequential Recommendation
- Evaluating LLM-based conversational recommenders with semantic and genre-aware metrics
- Integrating implicit desire fusion for proactive conversational recommendation systems with LLMs
Personalization · 个性化推荐 — 3 篇入选
头条:A Framework for Value-Aligned Modeling and Robust Experimentation — Google Discover 广告团队发现,转化预估模型(pConvs)的离线 Poisson Loss 和线上广告主价值(Delivered Value,每次点击的转化数乘广告主目标单次转化出价 tCPA 再求和)之间的相关系数只有 0.1,离线变好的模型上线后经常不涨。作者把损失里每次点击按 tCPA 加权,并给所有低流量实验配一个同样流量、重新训练的生产模型作对照。按 tCPA 加权后,离线损失衡量的就是广告主真正愿意付钱的那部分转化。离线与线上的相关系数升到 0.65,低流量与高流量实验的相关系数从 0.53 升到 0.93。
- A Framework for Value-Aligned Modeling and Robust Experimentation(深度解读)
- scikit-rec: A Unified, Extensible Recommendation Library with scikit-rec-agent, a Conversational Interface(深度解读)
- Generalized Position-Based Model: Rethinking Position Weights in Ranking Off-Policy Evaluation
AI Search · AI 搜索 — 10 篇入选
头条:The Fusion Tax: A Diagnostic Analysis of When Hybrid Retrieval Hurts — 混合检索把 BM25 和向量检索两路结果融合成一个排序,主流检索引擎和 RAG 框架的文档大多默认用倒数排名融合(RRF)。作者在 7 个 BEIR 数据集上逐条查询记账:表现最好的融合方式是向量分数占 0.7 权重的分数插值,它在每个数据集的平均 nDCG@10 上都赢过只用向量检索,却仍让 8–21% 的查询排得比只用向量检索更差。这些损失集中在只有一路命中相关文档的查询上,另一路此时只贡献噪声,融合会把已经排上来的相关文档往下压。把向量编码器从 22M 换到 596M 参数,这部分损失也没有消失。RRF 只在两路水平相当的数据集上才赢过纯向量检索,所以融合方式应该按目标语料上两路的实测差距来选。
- The Fusion Tax: A Diagnostic Analysis of When Hybrid Retrieval Hurts(深度解读)
- Chunk Recomputation under Real Document Revisions: Measuring Update Cost and Retrieval Quality of Fixed, Recursive, and Content-Defined Chunking(深度解读)
- Lean Finder V2: Semantic Search for Mathlib That Disambiguates User Queries(深度解读)
- CANOPY: Adaptive-Granularity Evidence Compression for Multimodal RAG
- 从 VDBBench 到 MMEB 榜首:阿里云 AI Search 从引擎到模型的全栈优化
- 大模型服务平台 TokenHub 联网搜索
- Joint Constraint Aggregation for Qualifier-Sensitive Retrieval-Augmented QA
- Compiling Autoregressive LLM Control into Policy-Exact Parallel Inference for Hierarchical RAG
- SurveyGraph: Generating Academic Surveys via Methods-Entity Knowledge Graphs
- Automated Enrichment Analysis with GraphRAG: Efficient Multi-Database ORA Extraction and Biological Interpretation
Infra · 系统与索引 — 8 篇入选
头条:llm-d 1.0, an open source stack for distributed inference — 智能体流量多步、共享长前缀、突发到达。放在普通轮询负载均衡后面,各副本会重复计算相同的提示,显存里堆满重复的 KV cache。Google 联合 Red Hat、IBM 等在 PyTorch Conference 发布的 llm-d 1.0,是一个跑在 Kubernetes 上的开源分布式推理栈,把一组 vLLM、SGLang 模型服务器组织成一个统一的服务池:路由器记录每个缓存前缀存在哪些副本上,结合负载和预测延迟选副本;显存满时把 KV cache 依次挪到主机内存和 SSD;预填充和解码拆成两个池。1.0 新增按优先级和公平性排队的流控,以及不中断会话的滚动升级。在 Ironwood TPU 上的编码智能体负载里,前缀感知路由加主机内存卸载让首 token 延迟中位数下降 85% 以上,总吞吐是普通 Kubernetes Service 的 2.4 倍。收益主要来自把同一会话的请求送回已经存着它上下文的副本,省掉重复的预填充。
- llm-d 1.0, an open source stack for distributed inference(深度解读)
- SweepANN: A GPU-Accelerated Large-K Approximate Vector Search for Drug Discovery
- llm-d 1.0 hardens open-source AI inference for production
- Amazon targets Trainium gains with new TorchNeuron developer tool
- FastLLM: a C++ LLM inference engine
- 向量检索数据库选型:Tair(Redis 企业版)内置向量引擎方案
- Scaling LLM Inference via Parallelism: Throughput-Latency Trade-off
- Introducing Superfluid: the local-AI native LLM server for multi-agent inference
近期日报
- 2026-10-24:Generative · Personalization · AI Search · Infra
- 2026-10-23:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-22:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-21:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-20:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-19:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-16:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-15:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-14:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-13:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-12:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-09:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-08:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-07:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-06:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-05:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-02:Generative · Personalization · AI Search · Agentic · Infra
- 2026-10-01:Generative · Personalization · AI Search · Agentic · Infra
- 2026-09-30:Generative · Personalization · AI Search · Agentic · Infra
- 2026-09-25:Generative · Personalization · AI Search · Agentic · Infra
- 2026-09-24:Generative · Personalization · AI Search · Agentic · Infra
- 2026-09-22:Generative · Personalization · AI Search · Agentic · Infra
- 2026-09-21:Generative · Personalization · AI Search · Agentic · Infra
- 2026-09-18:Generative · Personalization · AI Search · Agentic · Infra
- 2026-09-17:Generative · Personalization · AI Search · Agentic · Infra
- 2026-09-16:Generative · Personalization · AI Search · Agentic · Infra
- 2026-09-15:Generative · Personalization · AI Search · Agentic · Infra
- 2026-09-14:Generative · Personalization · AI Search · Agentic · Infra
- 2026-09-11:Generative · Personalization
- 2026-09-10:Generative · Personalization · AI Search · Agentic · Infra
更早的 141 期日报见 站点地图。