zzqdeco ~/blog

#Code Reading

共 16 篇 · 第 1 / 2 页
tags Deep Reading 33 Paper Reading 24 Code Reading 16 Computer Vision 12 Generative AI 5 Recommendation System 5 3D Generation 4 Agentic RL 3
还有 96 个标签 →
Deep Learning 3 Diffusion Models 3 Engineering 3 Image-to-3D 3 Knowledge Graph 3 Reinforcement Learning 3 Text-to-3D 3 3D Scene Generation 2 3D Scene Synthesis 2 Asynchronous RL 2 Computer Graphics 2 Graphics 2 NeRF 2 Optimization 2 RAG 2 Ranking 2 Systems 2 Two-Tower 2 Workflow 2 YouTube 2 3D Assets 1 3D Gaussian Splatting 1 3D Reconstruction 1 3D Scene Editing 1 3D Scene Modeling 1 3D Vision 1 Agent Memory 1 AI 1 AI Agents 1 Alibaba 1 ANN 1 Authorization 1 CAD Alignment 1 CAD Retrieval 1 Claude Code 1 Codex 1 Cold Start 1 Content Provenance 1 Context Engineering 1 Credit Assignment 1 CTR Prediction 1 Cypher 1 Database Internals 1 DeepSeek 1 Einops 1 Fourier Convolution 1 Generative Models 1 Generative Recommenders 1 Git 1 Graph Database 1 Graphiti 1 HSTU 1 Hybrid Search 1 Image Inpainting 1 Image Watermarking 1 Knowledge Augmented Generation 1 LLM Inference 1 LLM Post-Training 1 LLM Reliability 1 Long-Horizon RL 1 Loss Landscape 1 Machine Learning 1 Meta 1 Multi-Task Learning 1 Multimodal LLM 1 Multitask Learning 1 Neo4j 1 Notes 1 OpenFGA 1 OpenSPG 1 Pose Estimation 1 Python 1 Qwen 1 ReBAC 1 Rectified Flow 1 Referring Segmentation 1 Search 1 Search Engine 1 Security 1 Semi-Autoregressive Generation 1 Sequential Recommendation 1 Single-view Reconstruction 1 Speculative Decoding 1 Survey 1 System Design 1 Tensor Programming 1 Theory 1 Training Systems 1 Tutorial 1 User Interest Modeling 1 Vector Search 1 Vespa 1 Visual Grounding 1 Works 1 Zanzibar 1 Zep 1

2026

2026.10

  1. Claude Code 精读：工具证据、上下文工程与减少模型幻觉的设计

    基于独立架构分析与官方文档，精读 Claude Code 如何通过工具取证、编辑前置检查、上下文管理、任务验证和执行约束降低无依据推断与虚假完成风险

    Deep ReadingClaude CodeAI Agents +3
2026.08

  1. TrustMark 论文精读：任意分辨率图像水印、频谱损失与可重水印工作流

    精读 TrustMark 如何通过 MUNIT 嵌入器、ResNet50 提取器、可微扰动、焦点频率损失、分辨率缩放和 TrustMark-RM 实现高质量鲁棒图像水印

    Deep ReadingPaper ReadingImage Watermarking +4
2026.07

  1. SUFLECA 论文精读：大规模 NOC 特征学习、几何一致匹配与零样本 CAD 对齐

    精读 SUFLECA 如何通过 12 数据集 NOC 弱监督、DPT 几何特征、互为 Top-k 对应、各向异性尺度共识和 SupeRANSAC，实现高效的单图 CAD-to-image 9D 对齐

    Deep ReadingPaper Reading3D Vision +4

  2. Neo4j 源码精读：属性图、Cypher 查询引擎、事务存储与图数据库工程

    从 Neo4j 2026.06.0 Community 源码与 Cypher 论文双线精读属性图模型、查询编译执行、事务日志、页缓存、索引、Bolt 协议和 Community/Enterprise 边界

    Deep ReadingNeo4jGraph Database +4

  3. LaMa 论文精读：傅里叶卷积、大掩码修复与跨分辨率泛化

    精读 LaMa 如何通过 Fast Fourier Convolution、高感受野感知损失和大掩码训练，在单阶段图像修复中恢复周期结构并泛化到训练未见的高分辨率

    Deep ReadingPaper ReadingImage Inpainting +4

  4. Graphiti / Zep 论文精读：双时态知识图谱、Agent 长期记忆与混合检索

    从 Zep 论文与 Graphiti v0.29.2 源码双线精读三层记忆图、双时态事实、增量知识抽取、混合检索、重排策略和 Agent 长期记忆工程

    Deep ReadingPaper ReadingAgent Memory +4

  5. Diorama 论文精读：零样本单图 3D 场景建模、CAD 检索与语义布局优化

    精读 Diorama 如何通过开放世界感知、PlainRecon、场景图、CAD 检索、9-DoF 位姿估计和分阶段语义布局优化，从单张图像构建可编辑 3D 场景

    Deep ReadingPaper Reading3D Scene Modeling +4

  6. AgentRL 论文精读：全异步 Rollout-Training、跨策略采样与多任务 Agent 强化学习

    精读 AgentRL 如何通过全异步多轮训练、统一环境控制面、Cross-Policy Sampling、Task Advantage Normalization 和开源训练框架扩展多任务 Agentic RL

    Deep ReadingPaper ReadingReinforcement Learning +4

  7. OpenSPG KAG 论文精读：知识增强生成、逻辑形式推理与专业领域问答

    从 KAG 论文与 OpenSPG/KAG v0.8.0 源码双线精读 LLMFriSPG、知识-文本互索引、语义对齐、逻辑形式引导混合推理和专业知识服务工程实现

    Deep ReadingPaper ReadingKnowledge Graph +4

  8. Einops 论文精读：Einstein-like Pattern Language、张量形状语义与跨框架执行

    从 ICLR 2022 论文与 einops 0.8.2 源码双线精读 rearrange、reduce、repeat、pack、einsum、Pattern Parser、Transformation Recipe 与多后端执行机制

    Deep ReadingPaper ReadingPython +4