Appendix A — Glossary, English–Chinese
This glossary collects every term from the term tables at the end of the chapters: 732 of them, in alphabetical order. Each row gives the Chinese term that the Chinese edition of this book (《护栏之内》) uses, and the chapter whose table lists the term first. Click the chapter to go there.
The table is built from the chapters themselves each time the book is built. One English term has one Chinese term across the whole book. If two chapters ever disagree, the build stops until the chapter is fixed.
Words in brackets only say which sense is meant; they are not part of the term. Where one English word names two different things, the table keeps them apart: judge (Tom’s checking program) is not the debate judge of Chapter 27; a control in an experiment is not AI control; a layer of a network is not a safety layer. A chapter row with a slash, such as “test set / agreed set”, appears here as separate rows.
| English | 中文 | First chapter |
|---|---|---|
| 95% confidence interval (95% range) | 95% 置信区间(95% 区间) | Ch 15 |
| 95% interval | 95% 区间 | Prologue |
| 95% range | 95% 区间 | Ch 4 |
| A/B test | A/B 测试 | Ch 25 |
| ablation | 消融实验 | Ch 11 |
| access tier | 访问级别 | Ch 17 |
| account-level enforcement (account ban) | 账号级处置(封号) | Ch 10 |
| action | 行动 | Ch 24 |
| action-only monitor | 仅看行动的监控器 | Ch 24 |
| activations | 激活值 | Ch 9 |
| Activation Addition (ActAdd) | 激活加法 | Ch 21 |
| activation classifier | 激活分类器 | Ch 16 |
| activation steering | 激活引导 | Ch 21 |
| adaptive attack | 自适应攻击 | Ch 13 |
| adaptive attacker | 自适应攻击者 | Ch 11 |
| adaptive test | 自适应测试 | Ch 12 |
| adherence (to a constitution) | (对宪法的)遵循度 | Ch 5 |
| advantage | 优势值 | Ch 23 |
| adversarial example | 对抗样本 | Ch 19 |
| adversarial training | 对抗训练 | Ch 14 |
| agent | 智能体 | Ch 13 |
| agentic | 智能体式 | Ch 23 |
| agentic misalignment | 智能体未对齐 | Ch 26 |
| Agents Rule of Two | 智能体“三选二”规则 | Ch 13 |
| agreed set | 约定的测试集 | Ch 10 |
| agreement rate | 一致率 | Ch 5 |
| agreement subset (clear-cut cases) | 一致子集(明确案例) | Ch 24 |
| AI control | AI 控制 | Ch 27 |
| AI feedback | AI 反馈 | Ch 5 |
| AI Safety Level (ASL) | AI 安全等级(ASL) | Ch 10 |
| AI Security Institute (UK) | (英国)AI 安全研究所 | Ch 17 |
| alert threshold | 警戒阈值 | Ch 16 |
| aligned | 对齐的 | Ch 24 |
| aligned instruction | 一致的指令 | Ch 13 |
| alignment | 对齐 | Ch 2 |
| alignment audit | 对齐审计 | Ch 22 |
| alignment faking | 对齐伪装 | Ch 26 |
| annealing (training stage) | 退火(训练的最后阶段) | Ch 18 |
| answer key | 标准答案 | Ch 5 |
| approval fatigue | 审批疲劳 | Ch 13 |
| as rationalisation | 作为事后合理化的推理 | Ch 24 |
| assistant model (chat model) | 助手模型(对话模型) | Ch 1 |
| attack family | 攻击家族 | Ch 12 |
| attack success rate | 攻击成功率 | Ch 11 |
| attacker effort | 攻击者要花的力气(所需投入) | Ch 17 |
| attempt budget | 尝试预算 | Ch 12 |
| attention | 注意力(机制) | Ch 19 |
| attribution graph | 归因图 | Ch 20 |
| AUC | 曲线下面积 | Ch 12 |
| auditing agent | 审计智能体 | Ch 22 |
| auditing budget | 审计预算 | Ch 27 |
| auditing game | 审计游戏 | Ch 22 |
| AUPRC | 精确率–召回率曲线下面积(AUPRC) | Ch 8 |
| author concentration | 作者集中度 | Epilogue |
| automated auditor | 自动化审计模型 | Ch 18 |
| automated red teaming | 自动化红队测试 | Ch 17 |
| autonomous replication and adaptation | 自主复制与适应 | Ch 16 |
| auxiliary confidence loss | 辅助置信度损失 | Ch 27 |
| average over the message | 对整条消息取平均 | Ch 21 |
| averaged over runs (pooled) | 多次运行平均(合并) | Ch 15 |
| backdoor (hidden trigger) | 后门(隐藏触发器) | Ch 11 |
| base model | 基础模型 | Ch 1 |
| base model features viewed on a fine-tune | 在微调模型上看基础模型的特征 | Ch 19 |
| base rate | 基础比率 | Ch 8 |
| base64 | base64 编码 | Ch 13 |
| baseline | 基线 | Ch 10 |
| batch | 批次 | Ch 1 |
| (one version) beats (another) | (一个版本)胜过(另一个) | Ch 7 |
| benchmark | 基准测试 | Ch 15 |
| best-of-n | best-of-n(n 选一) | Ch 3 |
| best of N tries | N 次取最优 | Ch 14 |
| binary choice | 二选一的抉择 | Ch 26 |
| black-box | 黑盒 | Ch 14 |
| blind hand audit | 人工盲审 | Ch 24 |
| blue team | 蓝队 | Ch 27 |
| bootstrap | 自助法 | Ch 24 |
| Bradley–Terry model | Bradley–Terry 模型 | Ch 3 |
| bring your own policy | 自带规则 | Ch 8 |
| broad behaviour check | 广泛行为检查 | Ch 25 |
| buffer | 缓冲 | Ch 9 |
| bug bounty | 漏洞赏金计划 | Ch 9 |
| CAISI | (美国)AI 标准与创新中心 | Ch 17 |
| CAISSI (US) | 更名后的 CAISSI | Ch 17 |
| calibration half | 校准集(一半数据) | Ch 14 |
| cannot be ruled out | 无法排除 | Ch 16 |
| capabilities (provenance labels) | 能力标签(来源标签) | Ch 13 |
| capability threshold | 能力阈值 | Ch 16 |
| cascade | 级联 | Ch 9 |
| catch rate (sensitivity) | 检出率(灵敏度) | Ch 15 |
| causal test | 因果测试 | Ch 18 |
| CBRN (chemical, biological, radiological, nuclear) | 化生放核 | Ch 16 |
| ceiling (trained on correct labels) | 上限(用正确标签训练) | Ch 27 |
| centroid | 质心(平均向量) | Ch 22 |
| chain of command | 指挥链(权限层级) | Ch 5 |
| chain of thought | 思维链 | Ch 6 |
| chain-of-thought monitoring (reasoning monitor) | 思维链监控(推理监控) | Ch 24 |
| chat template | 对话模板 | Prologue |
| checkpoint | 检查点(模型存档) | Ch 5 |
| “Chekhov’s gun” | 契诃夫之枪 | Ch 26 |
| cipher | 密码(加密文字) | Ch 11 |
| circuit | 电路(也常译作“回路”) | Ch 20 |
| circuit breaker | 断路器 | Ch 14 |
| circuit tracing | 电路追踪 | Ch 20 |
| clamping (a feature) | 钳定(把特征固定在某个值) | Ch 19 |
| classifier | 分类器 | Ch 12 |
| closed model | 闭源模型 | Ch 11 |
| cluster bootstrap | 聚类自助法 | Epilogue |
| clustered standard errors | 聚类标准误 | Ch 15 |
| code check | 代码检查 | Ch 24 |
| code comment | 代码注释 | Ch 24 |
| coding score | 编程得分 | Prologue |
| Cohen’s kappa | 科恩 kappa 系数 | Ch 15 |
| coherence | 连贯性 | Ch 25 |
| coherent | 连贯的 | Ch 25 |
| cohort (of testers) | (测试人员)群组 | Ch 17 |
| collusion | 串通 | Ch 27 |
| competing objectives | 目标冲突 | Ch 6 |
| completeness score | 完整度得分 | Ch 20 |
| compliance gap | 服从差距 | Ch 26 |
| computer use | 计算机操作(智能体) | Ch 13 |
| computing power (compute) | 算力 | Ch 9 |
| confabulation | 虚构(编造的回答) | Ch 22 |
| confound | 混杂因素 | Ch 19 |
| confused deputy | 混淆代理人 | Ch 13 |
| confusion matrix | 混淆矩阵 | Ch 8 |
| confusion table | 混淆表 | Ch 15 |
| constitution (of a model) | 宪法(模型行为准则) | Ch 5 |
| Constitutional AI | 宪法式 AI | Ch 5 |
| Constitutional Classifiers | 宪法分类器 | Ch 9 |
| constrained fine-tuning | 约束微调 | Ch 11 |
| consultancy | 顾问模式 | Ch 27 |
| consultant | 顾问 | Ch 27 |
| contamination | 数据污染 | Ch 15 |
| content policy | 内容规范 | Ch 5 |
| context distillation | 上下文蒸馏 | Ch 6 |
| context summary | 上下文摘要 | Ch 26 |
| context window | 上下文窗口 | Ch 12 |
| contract term | 合同条款 | Ch 10 |
| contrast direction | 对比方向 | Ch 21 |
| contrast pair | 对比样本对 | Ch 26 |
| contrast question (contrast set) | 对照问题(对照集) | Ch 7 |
| control (in an experiment) | 对照(实验中的对照组) | Ch 18 |
| control evaluation | 控制评估 | Ch 27 |
| control protocol | 控制协议 | Ch 27 |
| control red team | 控制红队 | Ch 27 |
| controversial (label) | 有争议(标签) | Ch 8 |
| conversation monitor | 对话监控器 | Ch 14 |
| correlated failures | 相关失效 | Ch 10 |
| cosine similarity | 余弦相似度 | Ch 22 |
| covert action | 隐蔽行为 | Ch 18 |
| critic (feedback model) | 评审模型(反馈模型) | Ch 5 |
| Critical (capability level) | 关键(能力等级) | Ch 16 |
| Critical Capability Level (CCL) | 关键能力等级 | Ch 16 |
| critique and revise | 批评与修改 | Ch 5 |
| cross-layer transcoder | 跨层转码器 | Ch 20 |
| crowdworker | 众包工作者 | Ch 17 |
| cut-off (threshold) | 阈值 | Ch 21 |
| dangerous-capability evaluation | 危险能力评估 | Ch 16 |
| dashboard | 仪表盘 | Ch 25 |
| data augmentation | 数据增强 | Ch 9 |
| data filtering | 数据过滤 | Ch 18 |
| data log | 数据日志 | Ch 22 |
| datamarking | 数据标记 | Ch 13 |
| dataset | 数据集 | Ch 15 |
| day-to-day noise | 日常波动(噪声) | Ch 15 |
| dead feature | 死特征(从不激活的条目) | Ch 19 |
| debate | 辩论 | Ch 27 |
| debate judge | 辩论裁判 | Ch 27 |
| debater | 辩手 | Ch 27 |
| decoder | 解码器 | Ch 19 |
| defect | 倒戈 | Ch 26 |
| defection probe | 倒戈探针 | Ch 26 |
| defence in depth | 纵深防御 | Ch 10 |
| defer to trusted | 转交可信模型 | Ch 27 |
| definition gap | 定义缺口 | Ch 15 |
| degenerate (of a model’s text) | 退化(输出失去意义) | Ch 4 |
| deliberative alignment | 审慎对齐 | Ch 6 |
| delimiting | 分隔标记 | Ch 13 |
| design pattern | 设计模式 | Ch 13 |
| detection rate | 检出率 | Ch 12 |
| detection stack | 检测体系(多层检测) | Ch 14 |
| Developer | 开发者 | Ch 5 |
| dictionary entry | 词典条目 | Ch 19 |
| difference in differences | 双重差分 | Ch 15 |
| difference of means | 均值差 | Ch 21 |
| dimension | 维度 | Ch 19 |
| direct effect | 直接效应 | Ch 20 |
| Direct Preference Optimization (DPO) | 直接偏好优化(DPO) | Ch 4 |
| direct prompt injection (in some security guides) | 直接提示注入(部分安全指南的用法) | Ch 13 |
| direction (refusal direction) | 方向(拒绝方向) | Ch 11 |
| distilled away (reasoning) | (推理被)蒸馏掉 | Ch 26 |
| diversity (of test cases) | 多样性 | Ch 17 |
| drift | 偏移 | Ch 3 |
| drift gap | 偏移差 | Ch 4 |
| dual LLM pattern | 双模型模式 | Ch 13 |
| dual-use | 两用 | Ch 7 |
| early stopping | 提前停止 | Ch 4 |
| early-warning level | 预警水平 | Ch 16 |
| edge | 边 | Ch 20 |
| effectiveness | (测试用例的)有效性 | Ch 17 |
| elicitation | 能力激发(诱导) | Ch 16 |
| embedding | 嵌入(向量表示) | Ch 22 |
| embedding model | 嵌入模型 | Ch 22 |
| emergent misalignment | 涌现性未对齐 | Ch 22 |
| empty jailbreak | 空洞越狱 | Ch 14 |
| encoder | 编码器 | Ch 19 |
| encoding | 编码 | Ch 13 |
| engagement (testing contract) | 测试合作(委托测试) | Ch 17 |
| engineering trigger (if-then rule) | 工程触发条件(若-则规则) | Ch 16 |
| environment (RL) | 训练环境 | Ch 23 |
| environment realism | 环境真实性 | Ch 18 |
| episode | 回合 | Ch 23 |
| epoch | 轮次 | Ch 2 |
| error bar | 误差线 | Ch 15 |
| error node | 误差节点 | Ch 20 |
| evals (evaluations) | 评测 | Prologue |
| eval harness | 测试框架 | Ch 15 |
| evaluation awareness | 评测意识(察觉到自己正在被测试) | Ch 18 |
| exaggerated safety | 过度安全 | Ch 7 |
| exchange (a request and its reply) | 一轮交互(请求与回复) | Ch 9 |
| exchange classifier | 交互分类器 | Ch 9 |
| export controls | 出口管制 | Ch 10 |
| external red team | 外部红队 | Ch 17 |
| extremal Goodhart | 极端型古德哈特效应 | Ch 3 |
| fail | 未通过 | Ch 10 |
| faithfulness | 忠实性 | Ch 24 |
| fallback message | 兜底回复 | Ch 8 |
| false alarm (false positive) | 误报 | Ch 14 |
| false-alarm floor | 误报下限 | Epilogue |
| false-alarm rate | 误报率 | Ch 15 |
| false negative | 假阴性 | Ch 8 |
| false positive | 误报(假阳性) | Ch 8 |
| false-positive rate | 误报率 | Ch 8 |
| false refusal | 误拒 | Ch 7 |
| false refusal rate | 误拒率 | Ch 8 |
| feature | 特征 | Ch 19 |
| feature splitting | 特征分裂 | Ch 19 |
| fine-tune | 微调 | Ch 23 |
| fine-tuning | 微调 | Ch 2 |
| fishing (for the ending) | 套话(打探结局) | Ch 8 |
| fixed reply | 固定回复 | Ch 10 |
| flat refusal (hard refusal) | 生硬拒绝(硬拒绝) | Ch 7 |
| flawed code (the paper’s “backdoor”) | 有缺陷的代码(论文称“后门”) | Ch 27 |
| Fleiss’s kappa | 弗莱斯 kappa 系数 | Ch 17 |
| floor (smallest list that can pass) | 下限(能通过的最小样本量) | Ch 10 |
| fraction of variance unexplained (FVU) | 未解释方差比例 | Ch 19 |
| framing | 情境包装 | Ch 12 |
| free-form question | 开放式问题 | Ch 25 |
| fresh one-time exam | 全新的一次性考试 | Ch 15 |
| fresh problems | 全新题目 | Ch 23 |
| fresh test | 全新测试 | Ch 18 |
| frontier lab | 前沿实验室 | Ch 16 |
| frontier model | 前沿模型 | Ch 6 |
| Frontier Safety Roadmap | 前沿安全路线图 | Ch 16 |
| fuzzing | 模糊测试 | Ch 23 |
| gaming the tests | 钻测试的空子 | Ch 24 |
| garbled reply | 语无伦次的回答(乱码) | Ch 4 |
| generalise | 泛化 | Ch 3 |
| generator (model) | 生成模型 | Ch 9 |
| Goodhart’s law | 古德哈特定律 | Ch 3 |
| grader | 评分器 | Ch 7 |
| grader gaming | 钻评分漏洞 | Ch 15 |
| gradient descent | 梯度下降 | Ch 1 |
| gradient norm | 梯度范数 | Ch 4 |
| greedy decoding | 贪心解码 | Ch 8 |
| grey box | 灰盒 | Ch 17 |
| GRPO (Group Relative Policy Optimization) | 组相对策略优化 | Ch 23 |
| guard model | 防护模型 | Ch 8 |
| Guideline | 准则 | Ch 5 |
| hallucinated policy | 臆造的规则 | Ch 6 |
| hallucination | 幻觉 | Ch 1 |
| hard constraint | 硬性约束 | Ch 5 |
| hard label | 硬标签 | Ch 5 |
| hard line | 红线 | Ch 5 |
| harmful manipulation | 有害操纵 | Ch 16 |
| harmfulness rate | 有害率 | Ch 11 |
| harmless | 无害 | Ch 7 |
| head (classification head) | 分类头 | Ch 8 |
| hedge (a careful “maybe” answer) | 含糊其辞的说法 | Ch 9 |
| held-out model (trained without the secret) | 留出模型(训练时未见秘密) | Epilogue |
| held-out set | 留出集 | Ch 3 |
| held-out (sealed) test set | 留出(封存)测试集 | Ch 14 |
| helpful | 有用 | Ch 7 |
| helpful, honest, harmless (HHH) | 有用、诚实、无害 | Ch 2 |
| hidden objective | 隐藏目标 | Ch 22 |
| hidden tests (held-out tests) | 隐藏测试(留出测试) | Ch 23 |
| High | 高 | Ch 16 |
| honest way out | 诚实的退路 | Ch 26 |
| honeypot | 蜜罐(诱导性测试场景) | Ch 18 |
| human confirmation (human in the loop) | 人工确认(人在回路) | Ch 13 |
| human feedback | 人类反馈 | Ch 1 |
| identity preference optimisation (IPO) | 恒等偏好优化(IPO) | Ch 4 |
| imitation saliency | 模仿显著性 | Ch 27 |
| implied reward | 隐含奖励 | Ch 4 |
| in-band signalling | 带内信令 | Ch 13 |
| in context (probe reading a chat) | 在上下文中(探针读取整段对话) | Ch 21 |
| in-context learning | 上下文学习 | Ch 12 |
| in-sample | 样本内 | Ch 14 |
| incident response | 事故应对 | Ch 10 |
| independent | 独立失效 | Ch 10 |
| indirect helpfulness | 间接有用性 | Ch 7 |
| influence (in a graph) | 影响力(对输出的总影响) | Ch 20 |
| information asymmetry | 信息不对称 | Ch 27 |
| information hazard | 信息危害 | Ch 17 |
| inhibit | 抑制 | Ch 20 |
| inject (a feature) | 注入(特征) | Ch 20 |
| inoculation prompting | 接种提示 | Ch 23 |
| input classifier | 输入分类器 | Ch 8 |
| input filter | 输入过滤器 | Ch 11 |
| insider threat | 内部威胁 | Ch 26 |
| instruction hierarchy | 指令层级 | Ch 13 |
| intent | 意图 | Ch 7 |
| intercept | 截距 | Ch 12 |
| interference | 干扰 | Ch 19 |
| internal layer (of a model) | (模型的)内部层 | Ch 14 |
| intervention | 干预 | Ch 20 |
| inverse scaling | 逆向缩放 | Ch 13 |
| investigator agent | 调查智能体 | Ch 22 |
| jailbreak | 越狱 | Ch 6 |
| jailbreak severity scale | 越狱严重度等级 | Ch 10 |
| jailbroken model | 越狱模型 | Ch 25 |
| the judge (Tom’s checking program) | 判定程序 | Ch 1 |
| judge model | 评判模型 | Ch 6 |
| keyword search | 关键词检索 | Ch 22 |
| kill switch | 紧急开关 | Ch 10 |
| KL divergence | KL 散度 | Ch 3 |
| KL penalty (the leash) | KL 惩罚 | Ch 3 |
| KTO | KTO | Ch 4 |
| L0 | L0(每个词元激活的条目数) | Ch 19 |
| label | 标签 | Ch 8 |
| label noise | 标签噪声 | Ch 9 |
| labeller | 标注员 | Ch 3 |
| language model | 语言模型 | Ch 1 |
| last-token reading | 读最后一个词元 | Ch 21 |
| latent | 潜变量 | Ch 19 |
| launch-blocking | 阻止发布(的问题) | Ch 25 |
| launch stack | 上线防护组合 | Epilogue |
| layer | 层 | Ch 20 |
| leak or hint | 泄露或暗示 | Epilogue |
| leak pair | 泄露对 | Ch 5 |
| leak rate | 泄露率 | Prologue |
| leak-rate target | 泄露率目标 | Ch 10 |
| leash (pull back to the starting model) | 牵引绳(把模型拉回起步时的样子) | Ch 11 |
| leash strength | 牵引绳强度(β) | Ch 4 |
| least privilege | 最小权限 | Ch 13 |
| ledger | 账本 | Ch 10 |
| lexical overfitting | 词汇过拟合 | Ch 7 |
| linear probe | 线性探针 | Ch 21 |
| LLM (large language model) | 大语言模型 | Ch 13 |
| log-loss | 对数损失 | Ch 21 |
| log-odds | 对数几率 | Ch 5 |
| log-probability | 对数概率 | Ch 4 |
| logit | logit(未归一化的得分) | Ch 20 |
| LoRA (low-rank adaptation) | 低秩适配 | Ch 2 |
| loss | 损失 | Ch 1 |
| low-resource language | 低资源语言 | Ch 12 |
| malicious fine-tuning | 恶意微调 | Ch 18 |
| many-shot jailbreaking | 多样本越狱 | Ch 12 |
| matched false-positive rate | 匹配的误报率 | Ch 21 |
| matched operating point | 对齐的工作点 | Ch 9 |
| mean score | 平均分数 | Ch 14 |
| median | 中位数 | Ch 22 |
| mid-training | 中期训练 | Ch 23 |
| misaligned | 未对齐 | Ch 23 |
| misaligned instruction | 不一致的指令 | Ch 13 |
| “misaligned persona” feature | “未对齐人格”特征 | Ch 22 |
| misalignment | 未对齐 | Ch 23 |
| mismatched generalisation | 泛化错位 | Ch 11 |
| misuse | 滥用 | Ch 16 |
| MLP | MLP(多层感知机) | Ch 19 |
| the model being trained | 被训练的模型 | Ch 3 |
| model card | 模型卡 | Ch 8 |
| model diffing | 模型比对 | Ch 19 |
| model organism | 模式生物(用于研究的样本模型) | Ch 22 |
| model spec | 模型规范 | Ch 5 |
| modular pretraining | 模块化预训练 | Ch 18 |
| monitor | 监控器 | Ch 6 |
| monitorability | 可监控性 | Ch 24 |
| monitorability tax | 可监控性税 | Ch 24 |
| monitoring | 监控 | Ch 10 |
| monosemantic | 单义的 | Ch 19 |
| motivated reasoning | 动机性推理 | Ch 24 |
| n-gram | n 元语法 | Ch 15 |
| narrow fine-tuning | 窄范围微调 | Ch 25 |
| nats | 奈特(自然对数单位) | Ch 14 |
| natural language autoencoder | 自然语言自编码器 | Ch 19 |
| near-duplicate | 近似重复 | Ch 17 |
| necessity | 必要性 | Ch 24 |
| net change | 净变化 | Ch 10 |
| neuron | 神经元 | Ch 19 |
| next-token prediction | 下一个词元预测 | Ch 1 |
| node | 节点 | Ch 20 |
| not shown yet | 尚未证明 | Ch 10 |
| NP | NP | Ch 27 |
| obfuscated reward hacking | 隐蔽式奖励投机 | Ch 24 |
| offline evaluation | 离线评测 | Ch 25 |
| offline monitor | 离线监控器 | Ch 10 |
| one-time exam | 一次性考试 | Ch 18 |
| open (open-weight) model | 开放(开放权重)模型 | Ch 11 |
| open-weight model | 开放权重模型 | Ch 8 |
| OpenAI Model Spec | 《模型规范》(OpenAI) | Ch 13 |
| operating point | 工作点 | Ch 8 |
| opt-in alpha | 自愿参加的 alpha 测试 | Ch 25 |
| optimal stopping | 最优停止 | Ch 15 |
| optimisation pressure | 优化压力 | Ch 3 |
| optimised string | 优化字符串 | Ch 11 |
| optimiser | 优化器 | Ch 3 |
| ORPO | ORPO(DPO 的变体) | Ch 4 |
| out of scope | 超出范围 | Ch 26 |
| output classifier | 输出分类器 | Ch 8 |
| over-reach (of unlearning) | 过度删除 | Ch 18 |
| over-refusal | 过度拒绝 | Ch 2 |
| over-refusal pair | 过度拒绝对 | Ch 5 |
| overeager action | 越权操作(用户未要求的操作) | Ch 13 |
| overfitting | 过拟合 | Ch 4 |
| oversight | 监督 | Ch 27 |
| owner | 负责人 | Ch 17 |
| P(yes) | P(是) | Ch 24 |
| paired bootstrap | 配对自助法 | Ch 15 |
| paired comparison | 配对比较 | Ch 6 |
| paired range | 配对区间 | Ch 12 |
| paired result | 配对结果 | Ch 10 |
| paired test | 配对检验 | Ch 12 |
| parameters | 参数 | Ch 1 |
| parameterised query | 参数化查询 | Ch 13 |
| Pareto frontier | 帕累托前沿 | Ch 6 |
| patch | 补丁(修补) | Ch 10 |
| per-message filter | 逐条过滤器 | Ch 12 |
| percentage point | 百分点 | Ch 2 |
| percentile | 百分位数 | Ch 14 |
| performance gap recovered (PGR) | 性能差距弥补比例(PGR) | Ch 27 |
| perplexity | 困惑度 | Ch 12 |
| perplexity filter | 困惑度过滤器 | Ch 12 |
| persona | 人设 | Ch 2 |
| persona feature | 人格特征 | Ch 22 |
| persona vector | 人格向量 | Ch 21 |
| phase change | 相变 | Ch 19 |
| plan-then-execute | 先计划后执行 | Ch 13 |
| planning (a rhyme) | 规划(押韵词) | Ch 20 |
| policy (RL: the model being trained) | 策略(强化学习中被训练的模型) | Ch 24 |
| polysemantic | 多义的 | Ch 19 |
| position bias | 位置偏差 | Ch 5 |
| position paper | 立场论文 | Ch 24 |
| post-mortem | 事后复盘 | Ch 25 |
| post-training | 后训练 | Ch 1 |
| power law | 幂律 | Ch 12 |
| pre-registered questions | 预注册的问题 | Ch 25 |
| pre-registered (fixed in advance) threshold | 预先固定的阈值 | Ch 21 |
| precision | 精确率 | Ch 8 |
| precision at k | 前 k 项精确率 | Ch 22 |
| precision–recall curve | 精确率–召回率曲线 | Ch 8 |
| preference data | 偏好数据 | Ch 3 |
| preference pair (chosen / rejected answer) | 偏好对(被选回答 / 被拒回答) | Ch 3 |
| prefix | 前缀 | Ch 9 |
| Preparedness Framework | 准备度框架 | Ch 16 |
| preprint | 预印本 | Ch 11 |
| pretrained model | 预训练模型 | Ch 23 |
| pretraining | 预训练 | Ch 1 |
| pretraining leakage | 预训练渗漏 | Ch 27 |
| preventative steering | 预防性引导 | Ch 21 |
| principle | 原则 | Ch 5 |
| priority (when goals conflict) | 优先级(目标冲突时) | Ch 5 |
| privilege level | 权限级别 | Ch 13 |
| privileged | 特权模型 | Ch 13 |
| probe (linear probe) | 探针(线性探针) | Ch 9 |
| projection | 投影 | Ch 21 |
| proliferation (of attack variants) | 变体扩增(攻击变体) | Ch 14 |
| prompt injection (indirect prompt injection) | 提示注入(间接提示注入) | Ch 13 |
| prompt sandwiching | 提示夹层 | Ch 13 |
| propensity | 倾向性 | Ch 24 |
| provenance (of data) | (数据)来源记录 | Ch 25 |
| proxy | 代理指标 | Ch 3 |
| proxy test | 代理测试 | Ch 16 |
| pruning | 剪枝 | Ch 20 |
| PSPACE (complexity classes) | PSPACE(复杂度类) | Ch 27 |
| quarantined model | 隔离模型 | Ch 13 |
| random-direction control | 随机方向对照 | Ch 26 |
| random seed | 随机种子 | Ch 11 |
| rank correlation (Spearman) | 秩相关(斯皮尔曼) | Ch 7 |
| ransomware | 勒索软件 | Ch 10 |
| rapid response | 快速响应 | Ch 10 |
| raw input | 原始输入 | Prologue |
| re-alignment | 重新对齐 | Ch 25 |
| reading vector | 读取向量 | Ch 21 |
| real-time check | 实时检查 | Ch 10 |
| realism filter | 真实性过滤器 | Ch 18 |
| reasoning as computation | 作为计算的推理 | Ch 24 |
| reasoning model | 推理模型 | Ch 6 |
| reasoning text | 推理文本 | Ch 6 |
| recall | 召回率 | Ch 8 |
| recall at a 5% false-positive rate | 5% 误报率下的召回率 | Ch 24 |
| recall at a fixed false-positive rate | 固定误报率下的召回率 | Ch 14 |
| recall at k | 前 k 项召回率 | Ch 22 |
| reconstruction | 重建 | Ch 19 |
| recovery example | 恢复样本 | Ch 11 |
| red team | 红队 | Ch 2 |
| red-team model (attacker model) | 红队模型(攻击方模型) | Ch 17 |
| red-team prompt | 红队提示 | Ch 5 |
| red teaming | 红队测试 | Ch 17 |
| redaction | 删减(涂黑) | Ch 16 |
| reference model (starting model) | 参考模型(初始模型) | Ch 3 |
| reference model in capability tests | 参照模型 | Ch 16 |
| refusal opening | 拒绝开头 | Ch 11 |
| refusal with redirection | 拒绝并引导 | Ch 7 |
| regressional | 回归型 | Ch 3 |
| reinforcement learning (RL) | 强化学习 | Ch 3 |
| reinforcement learning from AI feedback (RLAIF) | 基于 AI 反馈的强化学习 | Ch 5 |
| reinforcement learning from human feedback (RLHF) | 基于人类反馈的强化学习 | Ch 3 |
| rejection sampling (best of N) | 拒绝采样(N 选优) | Ch 17 |
| ReLU | ReLU(线性整流函数,把负数变为零) | Ch 19 |
| remainder (what a layer misses) | 余数(某一层漏掉的部分) | Ch 10 |
| reminder at the point of action | 在行动的那一刻提醒 | Ch 26 |
| replacement model | 替代模型 | Ch 20 |
| replacement score | 替代得分 | Ch 20 |
| representation | 表征 | Ch 14 |
| Representation Engineering | 表征工程 | Ch 21 |
| resampling | 重采样 | Ch 27 |
| Research Category | 研究类别 | Ch 16 |
| residual stream | 残差流 | Ch 20 |
| Responsible Scaling Policy | 负责任扩展政策 | Ch 10 |
| retain example | 保留样本 | Ch 2 |
| retest date | 复测日期 | Ch 17 |
| reward (score) | 奖励(分数) | Ch 3 |
| reward gap | 奖励差 | Ch 4 |
| reward hacking | 奖励投机 | Ch 3 |
| reward model | 奖励模型 | Ch 3 |
| reward modelling | 奖励建模 | Ch 27 |
| reward over-optimisation | 奖励过度优化 | Ch 3 |
| reward signal | 奖励信号 | Ch 25 |
| reward tampering | 奖励篡改 | Ch 23 |
| Risk Report | 风险报告 | Ch 10 |
| robustness | 鲁棒性 | Ch 13 |
| ROC curve | ROC 曲线 | Ch 12 |
| role | 角色 | Ch 13 |
| role markers | 角色标记 | Ch 12 |
| role-play | 角色扮演 | Ch 11 |
| Root | 根 | Ch 5 |
| routing (flag for a second look) | 分流(标记待复查) | Ch 14 |
| rubric | 评分细则 | Ch 9 |
| rubric grader | 按评分细则打分的评分器 | Ch 14 |
| rule of three | 三法则 | Ch 10 |
| rulebook (allowed / banned content) | 规则手册(允许 / 禁止的内容) | Ch 9 |
| rules of engagement | 交战规则(测试约定) | Ch 17 |
| runbook (incident plan) | 事故应对预案 | Ch 10 |
| running maximum | 累计最大值 | Ch 9 |
| sabotage | 蓄意破坏 | Ch 16 |
| safe completion | 安全补全 | Ch 7 |
| safe-completion training | 安全补全训练 | Ch 7 |
| safeguard | 安全保障措施 | Ch 16 |
| Safety Advisory Group | 安全咨询小组 | Ch 16 |
| safety argument | 安全论证 | Ch 16 |
| safety-capability parity | 安全-能力对等 | Ch 11 |
| safety case | 安全论证 | Ch 16 |
| safety classifier | 安全分类器 | Ch 8 |
| safety framework | 安全框架 | Ch 16 |
| safety layer | (安全)防护层 | Ch 10 |
| safety margin | 安全余量 | Ch 8 |
| safety promise | 安全承诺 | Prologue |
| safety score | 安全分数 | Ch 27 |
| safety specification (spec) | 安全规范 | Ch 6 |
| safety training | 安全训练 | Ch 26 |
| sample | 样本 | Ch 15 |
| sampling seed | 采样随机种子 | Ch 15 |
| sandbagging | 藏拙(有意隐藏实力) | Ch 16 |
| sandbox | 沙箱 | Ch 13 |
| sanitisation | 输入净化 | Ch 12 |
| saturated (test) | 饱和(测试) | Ch 16 |
| scaffold | 脚手架(智能体外围程序) | Ch 16 |
| scalable oversight | 可扩展监督 | Ch 27 |
| scale | 规模 | Ch 11 |
| scheming | 图谋(暗中追求非预期目标) | Ch 18 |
| score | 分数 | Ch 8 |
| scorer | 评分器 | Ch 15 |
| scoring model | 打分模型 | Ch 14 |
| scratchpad | 草稿纸 | Ch 24 |
| sealed test set | 封存测试集 | Ch 11 |
| security rule (policy) | 安全规则(策略) | Ch 13 |
| seed | 随机种子 | Ch 23 |
| selected questions | 挑选的问题 | Ch 25 |
| self-play | 自我博弈 | Ch 17 |
| semantic search | 语义检索 | Ch 22 |
| severity | 严重程度 | Ch 17 |
| shadow use (shadow deployment) | 影子部署 | Ch 9 |
| shallow safety alignment | 浅层安全对齐 | Ch 11 |
| side effect | 副作用 | Ch 21 |
| simplest setting | 最宽松设置 | Ch 24 |
| SimPO | SimPO | Ch 4 |
| sleeper agent | 潜伏特工(潜伏型模型) | Ch 26 |
| slope | 斜率 | Ch 12 |
| slow escalation (multi-turn attack) | 多轮逐步升级 | Ch 12 |
| soft label | 软标签 | Ch 5 |
| solution contamination | 答案污染 | Ch 15 |
| solver | 求解器 | Ch 15 |
| sparse | 稀疏 | Ch 19 |
| sparse autoencoder (SAE) | 稀疏自编码器 | Ch 19 |
| sparse dictionary | 稀疏词典 | Ch 19 |
| special-casing the tests | 针对测试特判 | Ch 23 |
| specification | 规格(写下来的目标) | Ch 23 |
| specification gaming | 规格投机 | Ch 23 |
| spoiler classifier | 剧透分类器 | Ch 8 |
| spot check | 抽查 | Ch 25 |
| spotlighting | 聚光标记 | Ch 13 |
| SQL injection | SQL 注入 | Ch 13 |
| staged release | 分阶段发布 | Ch 25 |
| standard error | 标准误 | Ch 15 |
| state (of a model, at one point) | (模型在某处的)状态 | Ch 21 |
| static | 静态 | Ch 12 |
| static test | 静态测试 | Ch 17 |
| statistical power | 统计功效 | Ch 15 |
| statistically significant | 统计显著 | Ch 5 |
| steering | 激活引导(运行时修改激活值来改变行为) | Ch 19 |
| steering strength | 引导强度 | Ch 21 |
| step | 训练步 | Ch 2 |
| step size | 步长 | Ch 3 |
| stratified sample | 分层抽样 | Ch 15 |
| streaming | 流式(流式输出) | Ch 9 |
| streaming output classifier | 流式输出分类器 | Ch 9 |
| strength (of an entry) | 强度(条目的激活值) | Ch 19 |
| strong student | 强学生 | Ch 27 |
| success definition | 成功判定标准 | Ch 14 |
| Superficial Alignment Hypothesis | 表层对齐假说 | Ch 2 |
| supernode | 超节点 | Ch 20 |
| superposition | 叠加 | Ch 19 |
| supervised fine-tuning (SFT) | 监督微调 | Ch 2 |
| supply-chain attack | 供应链攻击 | Ch 26 |
| surprise (negative log-likelihood) | 意外度(负对数似然) | Ch 14 |
| suspicion score | 可疑度分数 | Ch 27 |
| Swiss cheese model | 瑞士奶酪模型 | Ch 10 |
| switch to a less capable model (Anthropic: fallback) | 切换到能力较弱的模型(Anthropic 称 fallback) | Ch 9 |
| sycophancy | 奉承 | Ch 21 |
| synthetic data | 合成数据 | Ch 8 |
| synthetic documents | 合成文档 | Ch 26 |
| synthetic training data | 合成训练数据 | Ch 5 |
| System | 系统 | Ch 5 |
| system card | 系统卡 | Ch 5 |
| system message | 系统消息 | Ch 13 |
| system prompt | 系统提示 | Prologue |
| tamper resistance | 抗篡改性 | Ch 18 |
| target | 目标答案 | Ch 15 |
| target reply | 目标回答 | Ch 2 |
| targeted attack success rate | 定向攻击成功率 | Ch 13 |
| taxonomy (of harms) | (危害)分类体系 | Ch 8 |
| teacher forcing | 教师强制(把已有文本喂回模型) | Ch 19 |
| temperature | 温度 | Prologue |
| test (unit test) | 测试(单元测试) | Ch 23 |
| test case | 测试用例 | Ch 17 |
| test harness | 测试框架 | Ch 23 |
| test set | 测试集 | Prologue |
| think tags | 思考标签(<think> / </think>) |
Ch 6 |
| thinking mode (on / off) | 思考模式(开 / 关) | Ch 6 |
| threat model | 威胁模型 | Ch 16 |
| threshold | 阈值 | Ch 7 |
| thumbs-down rating | 点踩评分 | Ch 25 |
| thumbs-up | 点赞 | Ch 25 |
| time horizon | 时间跨度 | Ch 23 |
| to game (a score) | 钻(分数的)空子 | Ch 23 |
| token | 词元 | Ch 1 |
| token budget | 词元预算 | Ch 15 |
| tool | 工具 | Ch 13 |
| tool result | 工具结果 | Ch 13 |
| top-activating examples | 激活最强的样例 | Ch 19 |
| top-k | top-k | Ch 1 |
| top of the 95% range (upper bound) | 95% 区间上限(上界) | Ch 10 |
| top-p | top-p | Ch 1 |
| TopK dictionary | TopK 词典(只保留最强的 k 个条目) | Ch 19 |
| “toxic persona” feature | “有毒人格”特征 | Ch 22 |
| toy simulation | 玩具模拟 | Ch 3 |
| Tracked | 追踪类别 | Ch 16 |
| Tracked Capability Level (TCL) | 追踪能力等级 | Ch 16 |
| trade-off | 取舍 | Ch 7 |
| trade-off curve | 取舍曲线 | Ch 7 |
| training-data search | 训练数据检索 | Ch 22 |
| training example | 训练样本 | Ch 2 |
| training objective | 训练目标 | Ch 11 |
| training step | 训练步 | Ch 4 |
| trait transfer | 特质转移 | Ch 25 |
| transcoder | 转码器 | Ch 19 |
| transcript | 运行记录(对话记录) | Ch 15 |
| transfer | 迁移 | Ch 12 |
| transfer attack | 迁移攻击 | Ch 14 |
| transformer | Transformer 架构 | Ch 20 |
| triage | 分级处理(按轻重缓急排序) | Ch 17 |
| trigger (of a behaviour) | 触发条件 | Ch 22 |
| trigger word | 触发词 | Ch 7 |
| true negative | 真阴性 | Ch 8 |
| true positive | 真阳性 | Ch 8 |
| true quality | 真实质量 | Ch 3 |
| trusted access | 可信访问 | Ch 10 |
| trusted editing | 可信编辑 | Ch 27 |
| trusted model | 可信模型 | Ch 27 |
| trusted monitoring | 可信监控 | Ch 27 |
| try | 尝试 | Ch 23 |
| two-hop reasoning | 两跳推理 | Ch 20 |
| underfit | 欠拟合 | Ch 4 |
| unfinished reply | 未完成的回复 | Ch 6 |
| universal jailbreak | 通用越狱 | Ch 9 |
| unlearning | 遗忘学习(机器遗忘) | Ch 18 |
| unpaired comparison | 非配对比较 | Ch 15 |
| untrusted data | 不可信数据 | Ch 13 |
| untrusted model | 不可信模型 | Ch 27 |
| untrusted monitoring | 不可信监控 | Ch 27 |
| usefulness | 效用 | Ch 27 |
| User | 用户 | Ch 5 |
| utility | 任务完成率(效用) | Ch 13 |
| validation set | 验证集 | Ch 6 |
| validity | 效度 | Ch 15 |
| value head | 价值头 | Ch 9 |
| verbalised awareness | 说出口的意识(在文字里明说) | Ch 18 |
| verdict: pass | 判定:通过 | Ch 10 |
| view (reasoning only / code only / both) | 视角(仅推理 / 仅代码 / 两者) | Ch 24 |
| violation rate | 违规率 | Ch 8 |
| vocabulary | 词表 | Ch 1 |
| warning phrase | 警示短语 | Ch 24 |
| weak supervisor | 弱监督者 | Ch 27 |
| weak-to-strong generalisation | 弱到强泛化 | Ch 27 |
| weights | 权重 | Ch 11 |
| white-box | 白盒 | Ch 14 |
| white-box access | 白盒访问 | Ch 22 |
| whole-conversation monitor | 整段对话监控器 | Ch 12 |
| Wilson interval | Wilson 区间 | Ch 15 |
| win chance (Bradley–Terry) | 胜出概率(Bradley–Terry) | Ch 4 |
| win rate | 胜率 | Ch 4 |
| window score | 窗口分数 | Ch 14 |
| word list | 词表 | Ch 24 |
| word overlap | 词重叠 | Ch 15 |
| worst-case testing | 最坏情况测试 | Ch 18 |
| writing chance (of an answer) | 写出概率(一字不差写出某个回答的概率) | Ch 4 |
| written rule (for a monitor) | 书面规则 | Ch 24 |
| written scope | 书面的任务范围 | Ch 26 |
| zero-shot | 零样本 | Ch 12 |
| zero-sum | 零和 | Ch 27 |
| zone of ambiguity | 模糊地带 | Ch 16 |