https://chenhuiyu.github.io/blog Notes on models, systems, and being human — in English and Chinese. en-SG Wed, 29 Jul 2026 00:00:00 GMT https://chenhuiyu.github.io/blog/generative-recommendation-05-p5-zh https://chenhuiyu.github.io/blog/generative-recommendation-05-p5-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation P5 把评分、下一项推荐、解释、评论理解与直接推荐统一成 text-to-text：同一个模型、同一种损失，只改变 personalized prompt。 https://chenhuiyu.github.io/blog/generative-recommendation-06-m6-rec-zh https://chenhuiyu.github.io/blog/generative-recommendation-06-m6-rec-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation M6-Rec 把统一推荐推进到真实工业链路：召回、CTR、解释、对话和内容创作共享语言底座，再用 option tuning、late interaction 与蒸馏控制成本。 https://chenhuiyu.github.io/blog/generative-recommendation-07-tallrec-zh https://chenhuiyu.github.io/blog/generative-recommendation-07-tallrec-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation TALLRec 用少量 recommendation instruction 与 LoRA 把 LLaMA 对齐到偏好预测，并用接近随机的零样本表现证明：语言知识不等于推荐知识。 https://chenhuiyu.github.io/blog/generative-recommendation-08-rella-zh https://chenhuiyu.github.io/blog/generative-recommendation-08-rella-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation ReLLa 发现 LLM 在远未用满 context window 时就无法利用更长行为序列，并用目标相关行为检索 SUBR 与混合样本微调 ReiT 提高信噪比。 https://chenhuiyu.github.io/blog/generative-recommendation-09-llara-zh https://chenhuiyu.github.io/blog/generative-recommendation-09-llara-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation LLaRA 把传统序列推荐器的 Item ID embedding 投影成行为 token，与商品标题 token 拼接，并通过 curriculum learning 让 LLM 同时利用世界知识与协同规律。 https://chenhuiyu.github.io/blog/generative-recommendation-10-dsi-zh https://chenhuiyu.github.io/blog/generative-recommendation-10-dsi-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation DSI 把文档写进模型参数，让同一个 Transformer 同时学习 document→ID 与 query→ID；检索从向量近邻搜索变成了受条件约束的 ID 生成。 https://chenhuiyu.github.io/blog/generative-recommendation-11-tiger-zh https://chenhuiyu.github.io/blog/generative-recommendation-11-tiger-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation TIGER 用 RQ-VAE 把商品内容压缩成粗到细的 Semantic ID，再让 T5 逐 token 生成下一件商品，为生成式推荐建立了经典范式。 https://chenhuiyu.github.io/blog/generative-recommendation-12-gptrec-zh https://chenhuiyu.github.io/blog/generative-recommendation-12-gptrec-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation GPTRec 对比一次打分的 Top-K 与逐件生成的 Next-K，并用 SVD 量化把每件商品表示成多个紧凑 token。 https://chenhuiyu.github.io/blog/generative-recommendation-13-lc-rec-zh https://chenhuiyu.github.io/blog/generative-recommendation-13-lc-rec-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation LC-Rec 用 Semantic ID 连接 LLaMA 与商品目录，再通过显式、隐式对齐任务把语言语义和协同语义写进同一模型。 https://chenhuiyu.github.io/blog/generative-recommendation-14-letter-zh https://chenhuiyu.github.io/blog/generative-recommendation-14-letter-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation LETTER 把层次语义、协同对齐与码本多样性写进同一个 tokenizer 目标，并用 ranking-guided generation 改善生成排序。 https://chenhuiyu.github.io/blog/generative-recommendation-15-etegrec-zh https://chenhuiyu.github.io/blog/generative-recommendation-15-etegrec-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation ETEGRec 用序列—商品与偏好—语义两类对齐损失连接 tokenizer 和生成器，再通过交替优化让商品 ID 与用户表示共同演化。 https://chenhuiyu.github.io/blog/generative-recommendation-16-hstu-zh https://chenhuiyu.github.io/blog/generative-recommendation-16-hstu-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation HSTU 把工业推荐重写为用户行为的 sequential transduction，并从注意力、采样、内核和服务四层解决长序列规模化。 https://chenhuiyu.github.io/blog/generative-recommendation-17-onerec-zh https://chenhuiyu.github.io/blog/generative-recommendation-17-onerec-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation OneRec 用 Balanced Semantic ID、Session-wise Generation、Sparse MoE 与迭代 DPO，把快手的召回—粗排—精排漏斗改写为一次列表生成。 https://chenhuiyu.github.io/blog/generative-recommendation-18-mtgr-zh https://chenhuiyu.github.io/blog/generative-recommendation-18-mtgr-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation MTGR 保留 DLRM 的候选交叉特征，把同一用户的多候选聚合成一个 HSTU 序列，并用 GLN 与动态遮罩兼顾规模化、效果与因果安全。 https://chenhuiyu.github.io/blog/generative-recommendation-19-onerec-think-zh https://chenhuiyu.github.io/blog/generative-recommendation-19-onerec-think-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation OneRec-Think 先将商品 code 对齐到语言空间，再用推荐 CoT 与 Rollout-Beam GRPO 激活推理，并以 Think-Ahead 把慢思考拆到离线。 https://chenhuiyu.github.io/blog/generative-recommendation-20-onereason-zh https://chenhuiyu.github.io/blog/generative-recommendation-20-onereason-zh Wed, 29 Jul 2026 00:00:00 GMT zh-CN Generative Recommendation OneReason 将推荐推理拆成 item perception 与 preference cognition，用四层任务、专才再统一的强化学习和快慢系统，回答“为什么 thinking mode 真的有效”。 https://chenhuiyu.github.io/blog/generative-recommendation-05-p5-en https://chenhuiyu.github.io/blog/generative-recommendation-05-p5-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation P5 rewrites rating, next-item prediction, explanation, review understanding, and direct recommendation as text-to-text: one model and one loss, with personalized prompts defining the task. https://chenhuiyu.github.io/blog/generative-recommendation-06-m6-rec-en https://chenhuiyu.github.io/blog/generative-recommendation-06-m6-rec-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation M6-Rec brings unified recommendation into a production stack: retrieval, CTR ranking, explanation, dialogue, and content creation share a language foundation, while option tuning, late interaction, and distillation control cost. https://chenhuiyu.github.io/blog/generative-recommendation-07-tallrec-en https://chenhuiyu.github.io/blog/generative-recommendation-07-tallrec-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation TALLRec aligns LLaMA with preference prediction using a small number of recommendation instructions and LoRA—and uses near-random zero-shot performance to show that language knowledge is not recommendation knowledge. https://chenhuiyu.github.io/blog/generative-recommendation-08-rella-en https://chenhuiyu.github.io/blog/generative-recommendation-08-rella-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation ReLLa finds that LLMs fail to benefit from longer behavior sequences far below their context limit, then raises signal-to-noise with target-conditioned behavior retrieval (SUBR) and mixed-sample tuning (ReiT). https://chenhuiyu.github.io/blog/generative-recommendation-09-llara-en https://chenhuiyu.github.io/blog/generative-recommendation-09-llara-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation LLaRA projects a sequential recommender’s Item ID embedding into a behavioral token, concatenates it with title tokens, and uses curriculum learning so the LLM can exploit both world knowledge and collaborative patterns. https://chenhuiyu.github.io/blog/generative-recommendation-10-dsi-en https://chenhuiyu.github.io/blog/generative-recommendation-10-dsi-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation DSI writes documents into model parameters and trains one Transformer on document-to-ID and query-to-ID mappings, turning retrieval into conditional ID generation. https://chenhuiyu.github.io/blog/generative-recommendation-11-tiger-en https://chenhuiyu.github.io/blog/generative-recommendation-11-tiger-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation TIGER compresses item content into coarse-to-fine Semantic IDs with an RQ-VAE, then trains T5 to generate the next item's address token by token. https://chenhuiyu.github.io/blog/generative-recommendation-12-gptrec-en https://chenhuiyu.github.io/blog/generative-recommendation-12-gptrec-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation GPTRec contrasts one-pass Top-K scoring with autoregressive Next-K list generation and quantizes SVD item coordinates into compact multi-token IDs. https://chenhuiyu.github.io/blog/generative-recommendation-13-lc-rec-en https://chenhuiyu.github.io/blog/generative-recommendation-13-lc-rec-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation LC-Rec connects LLaMA to a catalog through Semantic IDs, then uses explicit and implicit translation tasks to align language and collaborative semantics. https://chenhuiyu.github.io/blog/generative-recommendation-14-letter-en https://chenhuiyu.github.io/blog/generative-recommendation-14-letter-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation LETTER writes hierarchical semantics, collaborative alignment, and code diversity into one tokenizer objective, then adds ranking-guided generation. https://chenhuiyu.github.io/blog/generative-recommendation-15-etegrec-en https://chenhuiyu.github.io/blog/generative-recommendation-15-etegrec-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation ETEGRec connects the tokenizer and generator with sequence–item and preference–semantic alignment, then alternates optimization so item IDs and user states co-evolve. https://chenhuiyu.github.io/blog/generative-recommendation-16-hstu-en https://chenhuiyu.github.io/blog/generative-recommendation-16-hstu-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation HSTU reformulates industrial recommendation as sequential transduction over user actions and tackles long-sequence scaling across attention, sampling, kernels, and serving. https://chenhuiyu.github.io/blog/generative-recommendation-17-onerec-en https://chenhuiyu.github.io/blog/generative-recommendation-17-onerec-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation OneRec combines balanced Semantic IDs, session-wise generation, Sparse MoE, and iterative DPO to replace Kuaishou's retrieve–prerank–rank funnel with one list generator. https://chenhuiyu.github.io/blog/generative-recommendation-18-mtgr-en https://chenhuiyu.github.io/blog/generative-recommendation-18-mtgr-en Wed, 29 Jul 2026 00:00:00 GMT en Generative Recommendation MTGR preserves DLRM candidate crosses, aggregates many candidates into one HSTU user sequence, and uses GLN plus dynamic masks to combine scale, quality, and causal safety.