taac recordTransformer和Scaling Laws的框架下逐渐合流这篇文章试着用一条相对清晰的线把这个问题讲明白。一、先说结论推荐系统正在从“打分排序”走向“序列生成”传统推荐系统的主流程是工业界非常熟悉的多阶段级联架构先从海量物品中召回几千个候选再粗排到几百个再精排到几十个最后经过重排、多样性、规则策略等模块生成用户真正看到的列表这个流程非常强大也非常工程化。它最大的优点是稳定、可控、延迟低、容易拆分优化。每个阶段都有明确职责召回负责覆盖粗排负责效率精排负责精准重排负责业务约束和体验。但它也有天然问题。第一阶段之间会有信息损失。召回阶段没拿到的物品后面模型再强也排不出来。第二不同模块各自优化目标并不完全一致。召回看Recall粗排看 AUC精排看 CTR/CVR重排还要看多样性和策略规则最终用户体验其实是这些局部目标拼起来的结果第三模型规模继续变大时传统判别式模型很容易遇到瓶颈。尤其是推荐系统里有大量稀疏 ID 特征用户、物品、场景都在变化参数一大就容易过拟合不像 LLM 那样“越大越强”的规律那么自然。生成式推荐想做的事情更激进能不能用一个模型直接理解用户行为序列然后生成推荐结果也就是说把推荐系统从“多阶段候选集打分”改造成“端到端序列生成”这听起来很像 LLM。语言模型给定上文生成下一个 token生成式推荐给定用户历史行为生成下一个 item。一个续写文字一个续写兴趣。所以很多人说推荐系统正在迎来自己的“ChatGPT 时刻”。llm4rec 论文listOneRec Technical ReportOneRec-V2 Technical ReportOneRec: Unifying Retrieve and Rank with Generative Recommender and Preference AlignmentOneSearch: A Preliminary Exploration of the Unified End-to-End Generative Framework for E-commerce SearchQARM: Quantitative Alignment Multi-Modal Recommendation at KuaishouRankMixer: Scaling Up Ranking Models in Industrial RecommendersNext-User Retrieval: Enhancing Cold-Start Recommendations via Generative Next-User ModelingOneTrans: Unified Feature Interaction and Sequence Modeling with One Transformer in Industrial RecommenderMTGR: Industrial-Scale Generative Recommendation Framework in MeituanEGA-V2: An End-to-end Generative Framework for Industrial AdvertisingUniROM-One Model to Rank Them All: Unifying Online Advertising with End-to-End LearningSortGen-A Generative Re-ranking Model for List-level Multi-objective Optimization at TaobaoRecGPT: A Foundation Model for Sequential RecommendationCOBRA-Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense RepresentationsTowards Large-scale Generative RankingGoogle TIGER-Recommender Systems with Generative RetrievalGoogle Better Generalization with Semantic IDS: A Case Study in Ranking for RecommendationsMeta HSTU-Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative RecommendationsSnap GRID-Generative Recommendation with Semantic IDs: A Practitioner’s Handbook