English Glossary

This glossary keeps the English edition consistent. Prefer these terms when editing or extending the English version.

中文	English	Notes
大型语言模型 / LLM	large language model / LLM	Use "LLM" after first mention when natural.
续写	continuation	Prefer "continuation" over "completion" when explaining the model's core mechanism.
续写器	continuation engine	The book's central metaphor for base LLM behavior.
下一个 token 预测	next-token prediction	Keep "token" untranslated.
自回归生成	autoregressive generation	Use for the generation loop that appends one token at a time.
基座模型	base model	Avoid "foundation model" when referring specifically to the pretrained, non-instruction-tuned model.
对齐	alignment	Use consistently for SFT/RLHF/DPO/safety shaping.
预训练	pretraining	One word.
微调	fine-tuning	Hyphenated.
监督微调 / SFT	supervised fine-tuning / SFT	Spell out on first mention in a chapter.
偏好	preference	Used in RLHF/DPO contexts.
强化学习	reinforcement learning	Spell out unless paired with a standard acronym.
奖励模型	reward model	Use "RM" only after defining it.
幻觉	hallucination	Keep the established LLM term.
知识幻觉	knowledge hallucination	One of the three hallucination types in Chapter 7.
推理幻觉	reasoning hallucination	One of the three hallucination types in Chapter 7.
指令幻觉	instruction hallucination	One of the three hallucination types in Chapter 7.
忠实性	faithfulness	Use when discussing whether output is grounded in provided evidence or actual execution.
硬伤	hard limitations	Prefer this over "flaws" or "weaknesses" in chapter titles and summaries.
能力边界	capability boundaries	Use when discussing where LLMs are reliable or unreliable.
模式识别	pattern recognition	Core strength in Chapter 5.
类比	analogy	Use "analogical transfer" only when the sentence needs the process noun.
格式转换	format conversion	Chapter 5 and structured-output contexts.
信息提取	information extraction	Not just "extraction" in headings.
少样本学习	few-shot learning	Hyphenated.
上下文学习	in-context learning	Hyphenated.
思维链 / CoT	chain-of-thought / CoT	Spell out on first mention in a chapter.
推理模型	reasoning model	Use for o1/R1-style models and similar systems.
测试时计算	test-time compute	Hyphenated.
提示词 / Prompt	prompt	Lowercase except in headings or API names.
Prompt 工程	prompt engineering	Use sparingly when the original says it.
系统提示词	system prompt	Not "system instruction" unless referring to a vendor API label.
结构化输出	structured output	Chapter 9 and tool/function-call contexts.
函数调用	function calling	Keep vendor-neutral unless discussing a specific API.
受限解码	constrained decoding	Chapter 9 term.
知识注入	knowledge injection	Chapter 10 title term.
检索增强生成 / RAG	retrieval-augmented generation / RAG	Spell out on first mention in a chapter.
长上下文	long context	Use as a noun phrase.
嵌入 / Embedding	embedding	Lowercase except in headings or API names.
向量检索	vector retrieval	Use for retrieval over embeddings.
重排序 / 精排	reranking / reranker	"Reranker" for the model/component.
工具使用	tool use	Prefer over "tool usage."
智能体 / Agent	agent	Lowercase except in headings.
反思 / Reflection	reflection	Agent and prompting pattern.
多智能体	multi-agent	Hyphenated.
工作流	workflow	Distinguish from "agent."
评估	evaluation	Use "eval" only for code or the established engineering shorthand.
评估集	eval set	Matches the book's engineering tone.
回归测试	regression testing	Chapter 12 term.
可解释性 / Interpretability	interpretability	Lowercase except in headings.
机制可解释性	mechanistic interpretability	Chapter 13 term.
超位	superposition	Interpretability term.
稀疏自编码器 / SAE	sparse autoencoder / SAE	Spell out on first mention in a chapter.
回路	circuit	Use "circuits" for the research area.
特征 steering	feature steering	Keep "steering" because it is standard in this literature.
多模态	multimodal	One word.
全模态模型 / Omni Models	omni models	Lowercase in prose; title case in headings if matching source.
规模定律	scaling laws	Plural.
涌现能力	emergent abilities	Chapter 3 term.
过度训练	over-training	Hyphenated.
数据墙	data wall	Chapter 15 term.
开源 / 闭源	open source / closed source	No hyphen as noun phrase.