阅读指南与附录

参考资料与适用范围

阅读工具约 3.3 千字

以下出处支持正文中注明范围内的机制。中文讲解、推导和示例已在本地展开，链接用于追溯，不是必读的外部步骤。

如何区分三类材料

研究与官方文档说明公开事实；独立手算与构造案例用于解释机制；程序运行记录对应指定数据和断言。教学数值不代表商业模型参数，确定性测试不代表真实Agent成功率。产品与协议说明应与所列版本对应。

[TR] Attention Is All You Need

Transformer原始架构与注意力；现代Decoder变体另作说明。

论文 · 资料记录日期：2026-10-17

参考资料与适用范围

如何区分三类材料

[TR] Attention Is All You Need

[CHAT] Hugging Face：Chat templates

[BPE] Hugging Face：Byte-Pair Encoding tokenization

[UNIGRAM] Subword Regularization

[ROPE] RoFormer: Enhanced Transformer with Rotary Position Embedding

[GQA] GQA: Training Generalized Multi-Query Transformer Models

[CACHE] Hugging Face：KV cache strategies

[FLASH] FlashAttention: Fast and Memory-Efficient Exact Attention

[PAGED] Efficient Memory Management for LLM Serving with PagedAttention

[SPECULATIVE] Fast Inference from Transformers via Speculative Decoding

[RLHF] Training language models to follow instructions with human feedback

[DPO] Direct Preference Optimization

[R1] DeepSeek-R1

[LORA] LoRA: Low-Rank Adaptation of Large Language Models

[V3] DeepSeek-V3 Technical Report v2

[CODEX] Unrolling the Codex agent loop

[CODEX-SRC] Codex固定源码：工具路由与压缩

[TOOLS] OpenAI：Function calling

[COMPACT] OpenAI：Compaction

[PCACHE] OpenAI：Prompt caching

[CLAUDE] How Claude Code works

[COPILOT] About custom agents for GitHub Copilot CLI

[CONTEXT] Effective context engineering for AI agents

[CTXRETRIEVAL] Introducing Contextual Retrieval

[MCP] MCP Architecture overview

[MCP-LIFE] MCP Lifecycle，2025-11-25

[MCP-TOOLS] MCP Tools，2025-11-25

[SKILLS] Agent Skills specification

[SUBAGENTS] Claude Code：Subagents

[CLAUDEMEM] Claude Code：Memory

[COPILOTSKILL] GitHub Copilot：Agent Skills

[AGENTS] Building effective agents

[LONGHARNESS] Effective harnesses for long-running agents

[MULTI] How we built our multi-agent research system

[MEMGPT] MemGPT: Towards LLMs as Operating Systems

[COALA] Cognitive Architectures for Language Agents

[LONGMEM] LongMemEval

[DPR] Dense Passage Retrieval

[COLBERT] ColBERT

[HNSW] Efficient and robust approximate nearest neighbor search using HNSW

[BM25] Stanford IR Book：Okapi BM25

[RRF] Elasticsearch：Reciprocal rank fusion

[TREESITTER] Tree-sitter：Using parsers

[LSP] Language Server Protocol 3.17

[SWEBENCH] SWE-bench

[OTEL] OpenTelemetry：GenAI agent and framework spans

[GENERATION] Hugging Face：Generation strategies

[SPEC] OpenAI Model Spec，2025-12-18

[REASON] OpenAI：Reasoning models

[STRUCTURED] OpenAI：Structured outputs

[STREAM] OpenAI：Streaming responses

[SSE] WHATWG HTML：Server-sent events

[ADAMW] Decoupled Weight Decay Regularization

[CHINCHILLA] Training Compute-Optimal Large Language Models

[QLORA] QLoRA

[REACT] ReAct: Synergizing Reasoning and Acting

[LOST] Lost in the Middle

[HYDE] Precise Zero-Shot Dense Retrieval without Relevance Labels

[RAPTOR] RAPTOR

[GRAPHRAG] From Local to Global: A Graph RAG Approach

[SELFRAG] Self-RAG

[CRAG] Corrective Retrieval Augmented Generation

[TP] PyTorch：Tensor Parallel tutorial

[ZERO] ZeRO: Memory Optimizations Toward Training Trillion Parameter Models

[SRE] Google SRE：Monitoring Distributed Systems

[SLO] Google SRE Workbook：Implementing SLOs

[IDEMPOTENT] AWS Builders Library：Making retries safe with idempotent APIs

[SQLITE] SQLite：Isolation

[LATENCY] OpenAI：Latency optimization

[SANDBOX] Beyond permission prompts: Claude Code sandboxing

[A2A] Agent2Agent Protocol specification

[JUDGE] Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena

[HUMANEVAL] Evaluating Large Language Models Trained on Code

[TAU] τ-bench

[EVALS] Demystifying evals for AI agents

[REVIEW] Google Engineering Practices：The Standard of Code Review

[OWASP] OWASP LLM01：Prompt Injection

[HIERARCHY] The Instruction Hierarchy