LLM4Rec_wiki

KO | EN

gitlite — search

#javascript #python #ai #typescript #llm #golang #go #machine-learning #awesome #kubernetes #vue #pytorch #mcp #openai

(★ 11)

Automated LLM4Rec Wiki Project Driven by Large Language Models. LLM驱动维护的LLM4Rec知识库.

LLM4Rec_wiki 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# LLM4Rec Wiki

> 一个持久化、持续累积的知识库，专注于**大语言模型在推荐系统的应用 (LLM4Rec)**，基于 [karpathy/llm-wiki.md](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f) 模式构建，由Qwen Code调用qwen3.6-plus实现，使用阿里百炼Code Plan.

## 概述

本项目实现了一个由 LLM 维护的领域知识库，聚焦于**大语言模型在推荐系统的应用 (LLM4Rec)**。与传统的 RAG（每次查询都重新推导知识）不同，该 wiki **逐步构建并维护**一个结构化的、相互链接的知识库，随着时间推移不断累积。

## Summary

# LLM4Rec 知识库综合性总结

## 知识库统计

- **总页面数**：180
- **类别分布**：
  - 信息源（sources）：65 篇（占比最高，约 36%）
  - 模型（models）：31 篇
  - 方法（methods）：29 篇
  - 概念（concepts）：29 篇
  - 实体（entities）：16 篇
  - 总结（summary）：7 篇
  - 综合分析（synthesis）：3 篇

知识库以**论文源文献摘要**为主要驱动，模型、方法与概念三足鼎立，形成了"源文献→模型→方法→概念"的清晰知识链路。

---

## 核心覆盖领域

1. **生成式推荐范式（Generative Recommendation）**：知识库的核心主线，涵盖从 TIGER、OneRec、PLUM 到 OneRec-V2/Think 等模型，记录了从"双塔检索 + 排序"到"自回归生成 Token"的范式转变。

2. **Semantic ID 与离散标识符建模**：包括 FORGE、R3-VAE、ActionPiece、GRID 等工作，覆盖码本构建、码本崩溃缓解、SID 评估指标设计与冲突感知评估等关键议题。

3. **长序列建模与 Scaling Law**：以 HSTU、ULTRA-HSTU、LONGER、Wukong、Kunlun、GRAB 为代表，系统记录了推荐系统中算力-性能幂律关系的实证研究与硬件协同优化。

4. **多模态对齐与 LLM 用户理解**：包括 QARM V1/V2、LEMUR、LLM-Based User Personas 等，关注多模态特征与业务目标的定量对齐及 LLM 在线服务化路径。

5. **统一架构与工业部署**：OneTrans、OnePiece、HyFormer、RankMixer 等工作探索召回-排序-重排的统一骨干网络与工业级延迟约束下的可扩展架构设计。

---

## 关键方法与模型

1. **TIGER（Google, NeurIPS 2023）**：首个基于 Semantic ID 的生成式检索模型，开创了"自回归生成物品标识符"范式。

2. **OneRec 系列（快手）**：业界最早将检索与排序统一为端到端生成式模型；V2 版本引入惰性纯解码器，计算量降低 94%；OneRec-Think 进一步引入显式文内推理。

3. **HSTU / ULTRA-HSTU（Meta）**：分层序列转换单元，采用相对位置偏置与线性化注意力，将复杂度从 $O(L^2)$ 降至接近 $O(L)$，支撑万亿参数推荐模型。

4. **PLUM（Google/YouTube）**：Semantic ID 标记化 + 领域持续预训练 + 任务特定微调，将预训练 LLM 适配到工业级生成式推荐。

5. **FORGE（淘宝）**：基于 140 亿交互、2.5 亿商品的大规模 SID 构建框架，提出免训练代理评估指标。

6. **RankMixer + LONGER（字节跳动）**：硬件感知可扩展排序模型与长序列优化 Transformer，体现"主干可扩展化"的工业落地路径。

7. **MBGR（美团）**：通过业务感知语义 ID（BID）、多业务预测结构（MBP）与标签动态路由（LDR）解决"超级 App"多业务跷板效应。

8. **Wukong（Meta）**：纯 FM 堆叠架构 + 协同扩展策略，验证了无需复杂交叉网络也能实现推荐 Scaling Law。

---

## 工业实践与案例

知识库覆盖了**主流工业实体的完整实践图谱**：

- **快手**：OneRec 系列、QARM V1/V2、UA-SID 广告生成式推荐，在线 A/B 实测 App 停留时长提升 0.467%~0.741%
- **字节跳动/抖音**：LONGER、RankMixer、LEMUR 抖音搜索/广告全量部署
- **Google/YouTube**：TIGER、PLUM、LLM-Based User Personas 服务数十亿用户
- **阿里淘宝**：FORGE（交易额 +0.35%）、RecGPT/RecGPT-V2（CTR +2.98%, 新客探索率 +11.46%）
- **美团**：MBGR 外卖平台全量部署
- **腾讯**：HiGR（观看时长提升、推理速度 5 倍）、TAAC 算法挑战赛
- **Pinterest**：PinnerFormer 大规模生成式检索
- **Shopee**：OnePiece 核心搜索场景全量部署
- **百度**：GRAB 广告系统全量部署，广告总收入 +3.05%
- **京东**：OCP 正交约束投影，UCXR +12.97%、GMV +8.9%

---

## 研究前沿与开放问题

1. **生成式推荐的泛化性争议**：论文 26031980 指出 GR 模型的物品级泛化能力在 Token 层面常退化为"模式记忆"，GR 与传统 ID 模型的互补性是新兴研究方向。

2. **SID 评估的可靠性危机**：论文 26052533 揭示 SID 冲突在 4 数据集 5 分词器上比例高达 30.5%，Hit@10 被高估高达 103%，动摇了当前评估共识。

3. **显式推理与可控生成**：OneRec-Think、OxygenREC 的"快慢思考"双轨架构等工作正在探索将 CoT 推理引入推荐，但推理与延迟的平衡仍未解。

4. **稀疏-稠密协同扩展**：OCP、HyFormer 等工作开始关注嵌入层奇异值谱与稠密网络的协同扩展，但理论基础尚不完善。

5. **多业务/多目标统一建模**：MBGR 揭示的"跷板效应"与"表征混淆"在超级 App 场景仍是开放挑战，业务感知与统一表征的权衡有待深化。

---

## 未来建议的知识摄入方向

1. **强化学习与偏好对齐在生成式推荐中的应用**：当前知识库对 DPO、RSPO、Constrained RL 等方法记录较为零散，建议系统补充 RLHF/DPO 在 OneRec、RecGPT 等模型中的对比研究及奖励建模理论。

2. **冷启动与长尾泛化的系统性研究**：尽管 FORGE、R3-VAE 等涉及码本均衡，但针对 LLM4Rec 的冷启动专项方法（如检索增强生成 RAG-Rec、零样本提示工程）仍稀缺，建议补充。

3. **可解释性与对话式推荐**：InstructRec、P5 等早期工作已收录，但近期对话式推荐、Agent-based 推荐（如 RecAgent、智能体裁判）的最新进展覆盖不足。

4. **推荐系统的隐私、公平性与安全性**：当前知识库几乎完全聚焦性能与效率，缺乏对生成式推荐中数据隐私、推荐偏见、对抗鲁棒性等社会技术议题的覆盖。

5. **基础设施与训练系统优化**：FlashAttention 已收录，但 LLM4Rec 专用的分布式训练框架、KV 缓存复用、推理加速（如投机解码在生成式检索中的应用）仍是知识空白，建议优先补充工业级 Infra 论文。

6. **跨域与跨平台迁移学习**：现有工作多聚焦单一平台，但 LLM 的跨域迁移能力是其核心优势之一，建议补充跨域推荐基准（如 Amazon 多类目、跨 App 行为迁移）的最新研究。

## 快速开始
### 0. 创建独立环境
```bash
python -m venv venv_llm4rec_wiki
source venv_llm4rec_wiki/bin/activate
```

### 1. 安装依赖

```bash
pip install -r requirements.txt
```

### 2. 配置 API 密钥

```bash
# 获取你的密钥：https://bailian.console.aliyun.com/
export DASHSCOPE_API_KEY="sk-your-api-key-here or code-plan-api-key"
```

### 3. 使用 CLI

```bash
# 从 URL 抓取网页(Arxiv论文)并自动分析（推荐）
python tools/llm_wiki.py fetch "https://arxiv.org/abs/2307.06435"

# 导入新的源文档, 可以把网络上的碎片文档复制过来
python tools/llm_wiki.py ingest raw/sources/my_source.md

# 仅保存网页内容，不自动分析
python tools/llm_wiki.py fetch "https://example.com/article" --no-ingest

# 基于知识库，查询 wiki
python tools/llm_wiki.py query "生成式推荐和判别式推荐的本质区别？"

# 健康检查 wiki
python tools/llm_wiki.py lint

# 汇总当前 wiki 状态，并利用LLM生成Summary，并更新到README.md的Summary模块
python tools/llm_wiki.py summarize
```

## 项目结构

```
LLM4REC_wiki/
├── AGENTS.md              # Schema：规范与工作流
├── wiki/
│   ├── README.md          # Wiki 概览
│   ├── index.md           # 内容目录
│   ├── log.md             # 操作日志
│   ├── concepts/          # 核心理论
│   ├── methods/           # 算法
│   ├── models/            # 模型架构
│   ├── entities/          # 人物、数据集、平台
│   ├── synthesis/         # 跨主题分析
│   └── sources/           # 每个来源的摘要
├── raw/
│   ├── sources/           # 不可变的源文档
│   └── assets/            # 本地图片、图表
├── tools/
│   └── llm_wiki.py        # 集成百炼的 CLI 工具
├── scripts/               # 实用脚本
└── requirements.txt       # Python 依赖
```

## 当前内容

该 wiki 目前包含 **173 个页面**，涵盖：

### Concepts（29 页）
- [concepts/model_flops_utilization_mfu.md](wiki/concepts/model_flops_utilization_mfu.md)
- [concepts/slate_recommendation.md](wiki/concepts/slate_recommendation.md)
- [concepts/all_modality_gr.md](wiki/concepts/all_modality_gr.md)
- [concepts/hierarchical_planning_rec.md](wiki/concepts/hierarchical_planning_rec.md)
- [concepts/context_engineering_rec.md](wiki/concepts/context_engineering_rec.md)
- [concepts/llm4rec_overview.md](wiki/concepts/llm4rec_overview.md)
- [concepts/scaling_laws_recsys.md](wiki/concepts/scaling_laws_recsys.md)
- [concepts/continued_pretraining.md](wiki/concepts/continued_pretraining.md)
- [concepts/session_wise_generation.md](wiki/concepts/session_wise_generation.md)
- [concepts/representation_alignment.md](wiki/concepts/representation_alignment.md)
- [concepts/generative_retrieval.md](wiki/concepts/generative_retrieval.md)
- [concepts/evaluation_llm4rec.md](wiki/concepts/evaluation_llm4rec.md)
- [concepts/llm_alignment_and_optimization.md](wiki/concepts/llm_alignment_and_optimization.md)
- [concepts/prompt_engineering_rec.md](wiki/concepts/prompt_engineering_rec.md)
- [concepts/multimodal_recommendation.md](wiki/concepts/multimodal_recommendation.md)
- [concepts/heterogeneous_feature_interaction.md](wiki/concepts/heterogeneous_feature_interaction.md)
- [concepts/unified_transformer_backbone.md](wiki/concepts/unified_transformer_backbone.md)
- [concepts/collaborative_filtering.md](wiki/concepts/collaborative_filtering.md)
- [concepts/gsu_esu_paradigm.md](wiki/concepts/gsu_esu_paradigm.md)
- [concepts/semantic_id.md](wiki/concepts/semantic_id.md)
- [concepts/sequential_recommendation.md](wiki/concepts/sequential_recommendation.md)
- [concepts/intent_driven_recommendation.md](wiki/concepts/intent_driven_recommendation.md)
- [concepts/knowledge_enhanced_rec.md](wiki/concepts/knowledge_enhanced_rec.md)
- [concepts/memorization_vs_generalization.md](wiki/concepts/memorization_vs_generalization.md)
- [concepts/hybrid_generative_recommendation.md](wiki/concepts/hybrid_generative_recommendation.md)
- [concepts/recif_bench.md](wiki/concepts/recif_bench.md)
- [concepts/explicit_reasoning_rec.md](wiki/concepts/explicit_reasoning_rec.md)
- [concepts/end_to_end_multimodal_training.md](wiki/concepts/end_to_end_multimodal_training.md)
- [concepts/llm_user_persona.md](wiki/concepts/llm_user_persona.md)

### Methods（29 页）
- [methods/orthogonal_constrained_projection.md](wiki/methods/orthogonal_constrained_projection.md)
- [methods/quantitative_alignment.md](wiki/methods/quantitative_alignment.md)
- [methods/representation_alignment.md](wiki/methods/representation_alignment.md)
- [methods/synergistic_upscaling.md](wiki/methods/synergistic_upscaling.md)
- [methods/ua_sid.md](wiki/methods/ua_sid.md)
- [methods/adaptive_fusion_gr_id.md](wiki/methods/adaptive_fusion_gr_id.md)
- [methods/multi_objective_alignment.md](wiki/methods/multi_objective_alignment.md)
- [methods/asynchronous_llm_inference.md](wiki/methods/asynchronous_llm_inference.md)
- [methods/item_level_scoring_in_gr.md](wiki/methods/item_level_scoring_in_gr.md)
- [methods/iterative_preference_alignment.md](wiki/methods/iterative_preference_alignment.md)
- [methods/llm_distillation_for_rec.md](wiki/methods/llm_distillation_for_rec.md)
- [methods/rspo.md](wiki/methods/rspo.md)
- [methods/weighted_evaluation.md](wiki/methods/weighted_evaluation.md)
- [methods/reward_modeling_rec.md](wiki/methods/reward_modeling_rec.md)
- [methods/two_stage_training_rec.md](wiki/methods/two_stage_training_rec.md)
- [methods/rag_for_recsys.md](wiki/methods/rag_for_recsys.md)
- [methods/prompt_finetuning.md](wiki/methods/prompt_finetuning.md)
- [methods/multi_epoch_learning.md](wiki/methods/multi_epoch_learning.md)
- [methods/llm_as_generator.md](wiki/methods/llm_as_generator.md)
- [methods/collision_aware_evaluation.md](wiki/methods/collision_aware_evaluation.md)
- [methods/memory_bank_sequential_rep.md](wiki/methods/memory_bank_sequential_rep.md)
- [methods/human_llm_collaborative_evaluation.md](wiki/methods/human_llm_collaborative_evaluation.md)
- [methods/actionpiece_tokenization.md](wiki/methods/actionpiece_tokenization.md)
- [methods/lazy_ar.md](wiki/methods/lazy_ar.md)
- [methods/hybrid_generative_recommendation.md](wiki/methods/hybrid_generative_recommendation.md)
- [methods/sparse_attention_seq_rec.md](wiki/methods/sparse_attention_seq_rec.md)
- [methods/cama_attention.md](wiki/methods/cama_attention.md)
- [methods/llm_as_ranker.md](wiki/methods/llm_as_ranker.md)
- [methods/long_context_efficiency.md](wiki/methods/long_context_efficiency.md)

### Models（31 页）
- [models/2509_paper_25092090_FORGE_Forming_Semantic_Identifiers_for_Generative_Retrieval.md](wiki/models/2509_paper_25092090_FORGE_Forming_Semantic_Identifiers_for_Generative_Retrieval.md)
- [models/LEMUR.md](wiki/models/LEMUR.md)
- [models/OneRec.md](wiki/models/OneRec.md)
- [models/RecGPT.md](wiki/models/RecGPT.md)
- [models/LLMRank.md](wiki/models/LLMRank.md)
- [models/RankMixer.md](wiki/models/RankMixer.md)
- [models/P5.md](wiki/models/P5.md)
- [models/TIGER.md](wiki/models/TIGER.md)
- [models/InstructRec.md](wiki/models/InstructRec.md)
- [models/ULTRA_HSTU.md](wiki/models/ULTRA_HSTU.md)
- [models/SASRec.md](wiki/models/SASRec.md)
- [models/DHEN.md](wiki/models/DHEN.md)
- [models/LONGER.md](wiki/models/LONGER.md)
- [models/TALLRec.md](wiki/models/TALLRec.md)
- [models/HiGR.md](wiki/models/HiGR.md)
- [models/OneRec-Think.md](wiki/models/OneRec-Think.md)
- [models/Wukong.md](wiki/models/Wukong.md)
- [models/Hiformer.md](wiki/models/Hiformer.md)
- [models/PLUM.md](wiki/models/PLUM.md)
- [models/OnePiece.md](wiki/models/OnePiece.md)
- [models/qwen_series.md](wiki/models/qwen_series.md)
- [models/GRID.md](wiki/models/GRID.md)
- [models/Gryphon.md](wiki/models/Gryphon.md)
- [models/LLM_User_Personas.md](wiki/models/LLM_User_Personas.md)
- [models/HSTU.md](wiki/models/HSTU.md)
- [models/OneRec-V2.md](wiki/models/OneRec-V2.md)
- [models/ULTRA-HSTU.md](wiki/models/ULTRA-HSTU.md)
- [models/MBGR.md](wiki/models/MBGR.md)
- [models/GR4AD.md](wiki/models/GR4AD.md)
- [models/test_generative_rec_model.md](wiki/models/test_generative_rec_model.md)
- [models/QARM.md](wiki/models/QARM.md)

### Entities（16 页）
- [entities/movielens.md](wiki/entities/movielens.md)
- [entities/tencent.md](wiki/entities/tencent.md)
- [entities/baidu.md](wiki/entities/baidu.md)
- [entities/guorui_zhou.md](wiki/entities/guorui_zhou.md)
- [entities/shopee.md](wiki/entities/shopee.md)
- [entities/google_youtube.md](wiki/entities/google_youtube.md)
- [entities/jd.md](wiki/entities/jd.md)
- [entities/amazon_reviews.md](wiki/entities/amazon_reviews.md)
- [entities/pinterest.md](wiki/entities/pinterest.md)
- [entities/taac.md](wiki/entities/taac.md)
- [entities/kuaishou.md](wiki/entities/kuaishou.md)
- [entities/taobao.md](wiki/entities/taobao.md)
- [entities/meta.md](wiki/entities/meta.md)
- [entities/meituan.md](wiki/entities/meituan.md)
- [entities/bytedance.md](wiki/entities/bytedance.md)
- [entities/tencentgr_dataset.md](wiki/entities/tencentgr_dataset.md)

### Synthesis（3 页）
- [synthesis/traditional_vs_llm.md](wiki/synthesis/traditional_vs_llm.md)
- [synthesis/llm4rec_taxonomy.md](wiki/synthesis/llm4rec_taxonomy.md)
- [synthesis/lint_report_2026-04-08.md](wiki/synthesis/lint_report_2026-04-08.md)

### Sources（65 页）
- [sources/2603_paper_26031980_How_Well_Does_Generative_Recommendation_Generalize.md](wiki/sources/2603_paper_26031980_How_Well_Does_Generative_Recommendation_Generalize.md)
- [sources/2606_paper_26060860_Gryphon_A_Unified_Architecture_for_Semantic-ID_Generation_a.md](wiki/sources/2606_paper_26060860_Gryphon_A_Unified_Architecture_for_Semantic-ID_Generation_a.md)
- [sources/2510_paper_25102610_OneTrans_Unified_Feature_Interaction_and_Sequence_Modeling.md](wiki/sources/2510_paper_25102610_OneTrans_Unified_Feature_Interaction_and_Sequence_Modeling.md)
- [sources/2507_paper_25072222_Generative_Recommendation_with_Semantic_IDs_A_Practitioner'.md](wiki/sources/2507_paper_25072222_Generative_Recommendation_with_Semantic_IDs_A_Practitioner'.md)
- [sources/2509_paper_25092090_FORGE_Forming_Semantic_Identifiers_for_Generative_Retrieval.md](wiki/sources/2509_paper_25092090_FORGE_Forming_Semantic_Identifiers_for_Generative_Retrieval.md)
- [sources/2604_paper_26041144_R3-VAE_Reference_Vector-Guided_Rating_Residual_Quantization.md](wiki/sources/2604_paper_26041144_R3-VAE_Reference_Vector-Guided_Rating_Residual_Quantization.md)
- [sources/paper_4ddaf2_Recommender_Systems_with_Generative_Retrieval.md](wiki/sources/paper_4ddaf2_Recommender_Systems_with_Generative_Retrieval.md)
- [sources/2403_paper_24030254_Wukong_Towards_a_Scaling_Law_for_Large-Scale_Recommendation.md](wiki/sources/2403_paper_24030254_Wukong_Towards_a_Scaling_Law_for_Large-Scale_Recommendation.md)
- [sources/paper_c4a451_FORGE_Forming_Semantic_Identifiers_for_Generative_Retrieval.md](wiki/sources/paper_c4a451_FORGE_Forming_Semantic_Identifiers_for_Generative_Retrieval.md)
- [sources/2601_paper_26011268_HyFormer_Revisiting_the_Roles_of_Sequence_Modeling_and_Feat.md](wiki/sources/2601_paper_26011268_HyFormer_Revisiting_the_Roles_of_Sequence_Modeling_and_Feat.md)
- [sources/2411_paper_24111173_QARM_Quantitative_Alignment_Multi-Modal_Recommendation_at_K.md](wiki/sources/2411_paper_24111173_QARM_Quantitative_Alignment_Multi-Modal_Recommendation_at_K.md)
- [sources/2603_paper_26031869_OCP_Orthogonal_Constrained_Projection_for_Sparse_Scaling_in.md](wiki/sources/2603_paper_26031869_OCP_Orthogonal_Constrained_Projection_for_Sparse_Scaling_in.md)
- [sources/2407_paper_24072148_Breaking_the_Hourglass_Phenomenon_of_Residual_Quantization.md](wiki/sources/2407_paper_24072148_Breaking_the_Hourglass_Phenomenon_of_Residual_Quantization.md)
- [sources/2307_paper_23070643_A_Comprehensive_Overview_of_Large_Language_Models.md](wiki/sources/2307_paper_23070643_A_Comprehensive_Overview_of_Large_Language_Models.md)
- [sources/paper_QARM_V2_Quantitative_Alignment_Multi-Modal_Recommendation.md](wiki/sources/paper_QARM_V2_Quantitative_Alignment_Multi-Modal_Recommendation.md)
- [sources/2508_paper_25082090_OneRec-V2_Technical_Report.md](wiki/sources/2508_paper_25082090_OneRec-V2_Technical_Report.md)
- [sources/2604_paper_26040268_MBGR_Multi-Business_Prediction_for_Generative_Recommendatio.md](wiki/sources/2604_paper_26040268_MBGR_Multi-Business_Prediction_for_Generative_Recommendatio.md)
- [sources/paper_ad0dff_QARM_Quantitative_Alignment_Multi-Modal_Recommendation_at_K.md](wiki/sources/paper_ad0dff_QARM_Quantitative_Alignment_Multi-Modal_Recommendation_at_K.md)
- [sources/2602_paper_26022273_Generative_Recommendation_for_Large-Scale_Advertising.md](wiki/sources/2602_paper_26022273_Generative_Recommendation_for_Large-Scale_Advertising.md)
- [sources/2512_paper_25122238_OxygenREC_An_Instruction-Following_Generative_Framework_for.md](wiki/sources/2512_paper_25122238_OxygenREC_An_Instruction-Following_Generative_Framework_for.md)
- [sources/2606_paper_26061219_LLM-Based_User_Personas_for_Recommendations_at_Scale.md](wiki/sources/2606_paper_26061219_LLM-Based_User_Personas_for_Recommendations_at_Scale.md)
- [sources/2511_paper_25111096_LEMUR_Large_scale_End-to-end_MUltimodal_Recommendation.md](wiki/sources/2511_paper_25111096_LEMUR_Large_scale_End-to-end_MUltimodal_Recommendation.md)
- [sources/2512_paper_25121450_RecGPT-V2_Technical_Report.md](wiki/sources/2512_paper_25121450_RecGPT-V2_Technical_Report.md)
- [sources/2605_paper_26052533_How_Reliable_Are_Semantic-ID_Tokenizer_Comparisons_in_Genera.md](wiki/sources/2605_paper_26052533_How_Reliable_Are_Semantic-ID_Tokenizer_Comparisons_in_Genera.md)
- [sources/2502_paper_25021896_OneRec_Unifying_Retrieve_and_Rank_with_Generative_Recommend.md](wiki/sources/2502_paper_25021896_OneRec_Unifying_Retrieve_and_Rank_with_Generative_Recommend.md)
- [sources/2509_paper_25091809_OnePiece_Bringing_Context_Engineering_and_Reasoning_to_Indu.md](wiki/sources/2509_paper_25091809_OnePiece_Bringing_Context_Engineering_and_Reasoning_to_Indu.md)
- [sources/2205_paper_22050450_PinnerFormer_Sequence_Modeling_for_User_Representation_at_P.md](wiki/sources/2205_paper_22050450_PinnerFormer_Sequence_Modeling_for_User_Representation_at_P.md)
- [sources/2510_paper_25101163_OneRec-Think_In-Text_Reasoning_for_Generative_Recommendatio.md](wiki/sources/2510_paper_25101163_OneRec-Think_In-Text_Reasoning_for_Generative_Recommendatio.md)
- [sources/2512_paper_25122476_OpenOneRec_Technical_Report.md](wiki/sources/2512_paper_25122476_OpenOneRec_Technical_Report.md)
- [sources/2507_paper_25072287_RecGPT_Technical_Report.md](wiki/sources/2507_paper_25072287_RecGPT_Technical_Report.md)
- [sources/2206_paper_22060735_Rethinking_Reinforcement_Learning_for_Recommendation_A_Prom.md](wiki/sources/2206_paper_22060735_Rethinking_Reinforcement_Learning_for_Recommendation_A_Prom.md)
- [sources/2408_paper_24080458_FORGE_Force-Guided_Exploration_for_Robust_Contact-Rich_Mani.md](wiki/sources/2408_paper_24080458_FORGE_Force-Guided_Exploration_for_Robust_Contact-Rich_Mani.md)
- [sources/2311_paper_23110588_Hiformer_Heterogeneous_Feature_Interactions_Learning_with_T.md](wiki/sources/2311_paper_23110588_Hiformer_Heterogeneous_Feature_Interactions_Learning_with_T.md)
- [sources/2602_paper_26021001_Kunlun_Establishing_Scaling_Laws_for_Massive-Scale_Recommen.md](wiki/sources/2602_paper_26021001_Kunlun_Establishing_Scaling_Laws_for_Massive-Scale_Recommen.md)
- [sources/2305_paper_23051953_Multi-Epoch_Learning_for_Deep_Click-Through_Rate_Prediction.md](wiki/sources/2305_paper_23051953_Multi-Epoch_Learning_for_Deep_Click-Through_Rate_Prediction.md)
- [sources/paper_8edbf8_HiGR_Efficient_Generative_Slate_Recommendation_via_Hierarch.md](wiki/sources/paper_8edbf8_HiGR_Efficient_Generative_Slate_Recommendation_via_Hierarch.md)
- [sources/2602_paper_26020186_GRAB_An_LLM-Inspired_Sequence-First_Click-Through_Rate_Pred.md](wiki/sources/2602_paper_26020186_GRAB_An_LLM-Inspired_Sequence-First_Click-Through_Rate_Pred.md)
- [sources/2604_paper_26040497_Tencent_Advertising_Algorithm_Challenge_2025_All-Modality_G.md](wiki/sources/2604_paper_26040497_Tencent_Advertising_Algorithm_Challenge_2025_All-Modality_G.md)
- [sources/test_generative_rec.md](wiki/sources/test_generative_rec.md)
- [sources/2502_paper_25021358_ActionPiece_Contextually_Tokenizing_Action_Sequences_for_Ge.md](wiki/sources/2502_paper_25021358_ActionPiece_Contextually_Tokenizing_Action_Sequences_for_Ge.md)
- [sources/2402_paper_24021715_Actions_Speak_Louder_than_Words_Trillion-Parameter_Sequenti.md](wiki/sources/2402_paper_24021715_Actions_Speak_Louder_than_Words_Trillion-Parameter_Sequenti.md)
- [sources/2203_paper_22031101_DHEN_A_Deep_and_Hierarchical_Ensemble_Network_for_Large-Sca.md](wiki/sources/2203_paper_22031101_DHEN_A_Deep_and_Hierarchical_Ensemble_Network_for_Large-Sca.md)
- [sources/2510_paper_25102715_A_Survey_on_Generative_Recommendation_Data,_Model,_and_Task.md](wiki/sources/2510_paper_25102715_A_Survey_on_Generative_Recommendation_Data,_Model,_and_Task.md)
- [sources/1808_paper_18080978_Self-Attentive_Sequential_Recommendation.md](wiki/sources/1808_paper_18080978_Self-Attentive_Sequential_Recommendation.md)
- [sources/2205_paper_22051413_FlashAttention_Fast_and_Memory-Efficient_Exact_Attention_wi.md](wiki/sources/2205_paper_22051413_FlashAttention_Fast_and_Memory-Efficient_Exact_Attention_wi.md)
- [sources/paper_1b102d_QARM_V2_Quantitative_Alignment_Multi-Modal_Recommendation_f.md](wiki/sources/paper_1b102d_QARM_V2_Quantitative_Alignment_Multi-Modal_Recommendation_f.md)
- [sources/2306_paper_23060812_Better_Generalization_with_Semantic_IDs_A_Case_Study_in_Ran.md](wiki/sources/2306_paper_23060812_Better_Generalization_with_Semantic_IDs_A_Case_Study_in_Ran.md)
- [sources/2510_paper_25102127_Pctx_Tokenizing_Personalized_Context_for_Generative_Recomme.md](wiki/sources/2510_paper_25102127_Pctx_Tokenizing_Personalized_Context_for_Generative_Recomme.md)
- [sources/rankmixer_to_oneranker.md](wiki/sources/rankmixer_to_oneranker.md)
- [sources/2006_paper_20060577_Self-Supervised_Reinforcement_Learning_for_Recommender_Syste.md](wiki/sources/2006_paper_20060577_Self-Supervised_Reinforcement_Learning_for_Recommender_Syste.md)
- [sources/2602_paper_26020656_TokenMixer-Large_Scaling_Up_Large_Ranking_Models_in_Industr.md](wiki/sources/2602_paper_26020656_TokenMixer-Large_Scaling_Up_Large_Ranking_Models_in_Industr.md)
- [sources/2511_paper_25110607_Make_It_Long,_Keep_It_Fast_End-to-End_10k-Sequence_Modeling.md](wiki/sources/2511_paper_25110607_Make_It_Long,_Keep_It_Fast_End-to-End_10k-Sequence_Modeling.md)
- [sources/2602_paper_26021698_Bending_the_Scaling_Law_Curve_in_Large-Scale_Recommendation.md](wiki/sources/2602_paper_26021698_Bending_the_Scaling_Law_Curve_in_Large-Scale_Recommendation.md)
- [sources/2411_paper_24111881_Unifying_Generative_and_Dense_Retrieval_for_Sequential_Recom.md](wiki/sources/2411_paper_24111881_Unifying_Generative_and_Dense_Retrieval_for_Sequential_Recom.md)
- [sources/paper_81ec38_PLUM_Adapting_Pre-trained_Language_Models_for_Industrial-sc.md](wiki/sources/paper_81ec38_PLUM_Adapting_Pre-trained_Language_Models_for_Industrial-sc.md)
- [sources/2404_paper_24040057_A_Review_of_Modern_Recommender_Systems_Using_Generative_Mode.md](wiki/sources/2404_paper_24040057_A_Review_of_Modern_Recommender_Systems_Using_Generative_Mode.md)
- [sources/2511_paper_25111013_GPR_Towards_a_Generative_Pre-trained_One-Model_Paradigm_for.md](wiki/sources/2511_paper_25111013_GPR_Towards_a_Generative_Pre-trained_One-Model_Paradigm_for.md)
- [sources/2502_paper_25021647_Unified_Semantic_and_ID_Representation_Learning_for_Deep_Rec.md](wiki/sources/2502_paper_25021647_Unified_Semantic_and_ID_Representation_Learning_for_Deep_Rec.md)
- [sources/2505_paper_25050442_LONGER_Scaling_Up_Long_Sequence_Modeling_in_Industrial_Reco.md](wiki/sources/2505_paper_25050442_LONGER_Scaling_Up_Long_Sequence_Modeling_in_Industrial_Reco.md)
- [sources/2604_paper_26041373_TokenFormer_Unify_the_Multi-Field_and_Sequential_Recommenda.md](wiki/sources/2604_paper_26041373_TokenFormer_Unify_the_Multi-Field_and_Sequential_Recommenda.md)
- [sources/2507_paper_25071555_RankMixer_Scaling_Up_Ranking_Models_in_Industrial_Recommend.md](wiki/sources/2507_paper_25071555_RankMixer_Scaling_Up_Ranking_Models_in_Industrial_Recommend.md)
- [sources/paper_2305_05065_Generative_Retrieval_RecSys.md](wiki/sources/paper_2305_05065_Generative_Retrieval_RecSys.md)
- [sources/2602_paper_26020855_QARM_V2_Quantitative_Alignment_Multi-Modal_Recommendation_f.md](wiki/sources/2602_paper_26020855_QARM_V2_Quantitative_Alignment_Multi-Modal_Recommendation_f.md)
- [sources/paper_c33d89_Farewell_to_Item_IDs_Unlocking_the_Scaling_Potential_of_Lar.md](wiki/sources/paper_c33d89_Farewell_to_Item_IDs_Unlocking_the_Scaling_Potential_of_Lar.md)
- [sources/2504_paper_25041050_PinRec_Outcome-Conditioned,_Multi-Token_Generative_Retrieva.md](wiki/sources/2504_paper_25041050_PinRec_Outcome-Conditioned,_Multi-Token_Generative_Retrieva.md)

## 设计理念

基于 [Karpathy 的 LLM Wiki 模式](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f)：

> "Wiki 是一个持久化、持续累积的产物。交叉引用已经存在。矛盾之处已经被标记。综合内容已经反映了你阅读过的所有内容。"

### 三个层次
1. **源文档（Raw Sources）** — 不可变的输入文档
2. **Wiki** — LLM 生成的、相互链接的 Markdown 文件
3. **Schema** — 给 LLM 的指令（AGENTS.md）

### 三项操作
1. **Ingest（导入）** — 添加源文档 → LLM 更新 wiki
2. **Query（查询）** — 提出问题 → LLM 从 wiki 中综合生成答案
3. **Lint（检查）** — 健康检查 → LLM 发现矛盾、空白和孤立页面

## LLM 后端

- **平台**：阿里云百炼（Alibaba Cloud Bailian）
- **API 类型**：OpenAI 兼容 SDK（同时支持标准密钥和 Coding Plan 密钥）
- **Coding Plan 密钥**（`sk-sp-*`）：`qwen3.5-plus`（默认）、`qwen3-coder-plus`、`kimi-k2.5`、`glm-5`
- **标准密钥**（`sk-*`）：`qwen-max`（默认）、`qwen-plus`、`qwen-turbo`
- **自动检测**：CLI 会自动检测你的密钥类型并配置正确的端点

## 添加新源文档

### 方式 A：从 URL 抓取（推荐）
```bash
# 自动：抓取 → 保存 → 分析 → 更新知识库
python tools/llm_wiki.py fetch "https://arxiv.org/abs/2307.06435"

# 仅保存源文件，不自动分析
python tools/llm_wiki.py fetch "https://example.com/article" --no-ingest
```

### 方式 B：手动添加本地文件
1. 将源文件放置于 `raw/sources/`
2. 运行：`python tools/llm_wiki.py ingest raw/sources/your_file.md`
3. 审查建议的 wiki 变更
4. 确认应用变更

## 环境变量

| 变量 | 说明 | 默认值 |
|----------|-------------|---------|
| `DASHSCOPE_API_KEY` | 你的百炼 API 密钥 | （必填） |
| `BAILIAN_MODEL` | 主模型名称 | `qwen-max` |
| `WIKI_AUTO_CONFIRM` | 自动应用 wiki 变更 | `false` |

## 许可证

本知识库用于研究和教育目的。

## 致谢

- Karpathy 提出的 LLM Wiki 模式
- 阿里云百炼提供的 LLM 基础设施
- LLM4Rec 研究社区

LLM4Rec_wiki

LLM4Rec_wiki 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

LLM4Rec_wiki

LLM4Rec_wiki 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README