延伸阅读与来源索引

一本书不可能替读者读完所有论文,也不应让几百条参考文献变成一堵没有入口的墙。本附录的作用,是把全书已经引用的原始论文、教材、综述、评测和标准重新组织成可以执行的阅读路线

这里不是参考文献表的重复版本。文末参考文献回答“本书用了哪些来源”,本附录回答另外三个问题:

  1. 想把某一主题继续读深,应先读什么、后读什么?
  2. 一份材料适合用来理解历史、学习机制、复现结果,还是判断现实影响?
  3. 面对不断更新的模型、产品和政策,怎样保留可以复查的阅读记录?

读者不必从头读到尾。先选择自己的目标,再进入相应路线即可。

先选择阅读深度

主线读者:把概念连成一条路

目标是理解 AI 为什么经历规则、学习、深度网络、基础模型和智能体这些变化。每个主题只需完成三步:

  1. 先读本书对应章节,建立问题和词汇;
  2. 再读一篇历史材料、综述或教材章节;
  3. 最后读一篇标志性论文的摘要、引言、图表和局限。

主线读者不必逐式推导,也不必复现全部实验。能够回答“它解决了什么旧问题、依赖什么条件、还不能做什么”,就已经读到了关键处。

工程读者:追踪输入、输出和训练目标

除了主线问题,还要追踪:数据怎样进入系统,张量怎样变形,损失函数优化什么,训练和部署各需要什么资源,评测协议是否能支持论文的结论。建议把必要的数学准备放在手边,并给每篇论文画出一条最小数据流。

工程阅读的最低产物不是“我看懂了”,而是一页说明:

  • 输入、输出与主要张量形状;
  • 模型结构和目标函数;
  • 数据来源与切分方式;
  • 训练预算、推理预算和外部工具;
  • 主实验、消融实验与已知失败模式。

复现读者:把论文结论变成可失败的实验

复现不是运行作者仓库后得到一张相似截图。先选择要复现的最小结论,例如“残差连接使更深网络更容易优化”,而不是“重新训练完整模型”。随后固定版本、随机种子、数据切分、硬件、运行次数和评价指标,并预先写下什么结果会使复现失败。

对大模型和机器人系统,原论文规模通常超出个人预算。此时可以复现机制,不冒充复现规模:使用较小模型验证因果掩码、缩放点积注意力、检索增强或闭环控制的作用;不能由小实验推出原系统的绝对能力。

不同来源承担不同证据任务

  • 原始论文最适合确认作者提出了什么方法、怎样实验和如何限定结论。它不自动代表方法已经被独立复现。
  • 教材适合补齐共同语言、推导和成熟方法之间的关系。教材出版较慢,不适合确认最新产品状态。
  • 综述与历史著作适合建立全景和争议线索。它们的叙事选择仍需回到一手材料核对。
  • 技术报告与系统卡适合了解模型、数据、算力、评测和部署约束,但开发者也是利益相关方。
  • 基准及其数据说明定义“分数究竟测了什么”。只看排行榜而不看任务样本和协议,无法解释排名。
  • 独立复现与对抗测试能检查结果是否依赖作者代码、特定提示或隐藏条件。一次复现失败也可能来自实现差异,需要说明偏差来源。
  • 标准、法规和治理框架定义风险过程、责任和适用范围。提案、指南、标准、已通过法规和已生效规则不是同一种状态。
  • 部署研究最接近真实影响,却会受到组织流程、用户选择和时间变化的共同影响。

判断一项历史事实,优先找同期论文、档案或当事人的原始记录;学习一个成熟算法,教材往往比最早论文更清楚;判断一项当代能力,则要把系统报告、基准、独立测试和部署条件放在一起。来源没有抽象的“最高等级”,只有是否适合当前问题。

路线一:智能的观念、逻辑与计算

先读第 1 章第 4 章,再用 Nilsson 的 AI 史建立研究路线之间的联系。(Nilsson 2010年) 若想贴近原始材料,可以依次抽读 Lovelace 对分析机的注释、Boole 的逻辑代数、Church 与 Turing 对可计算性的形式化、Shannon 的继电器电路与信息论,以及 Wiener 的控制论。(Lovelace 1843年; Boole 1854年; Church 1936年; Turing 1937年; Shannon 1938年, 1948年; Wiener 1948年)

最后对读 Turing 1950 年的论文与达特茅斯提案。前者讨论“机器能否思考”应怎样转化为可讨论的问题,后者展示一个新研究领域怎样为自己划定议程。(Turing 1950年; McCarthy 等 1955年) 阅读时特别标记:哪些句子是技术结果,哪些是研究计划,哪些只是面向未来的判断。

推荐产物:画一条从“可形式化规则”到“可计算过程”再到“学习系统”的概念链,并为每个箭头写出一个它没有解决的问题。

路线二:经典机器学习与强化学习

先读第 8 章第 12 章。Mitchell 适合建立假设空间、泛化和归纳偏置的基本语言;Hastie 等与 Bishop 更适合继续学习统计学习、概率模型、正则化和模型选择。(Mitchell 1997年; Hastie 等 2009年; Bishop 2006年)

随后把算法放回原始转折:感知机、支持向量机、随机森林和提升方法分别展示了线性学习、间隔、集成与逐步纠错的不同思路。(Rosenblatt 1958年; Cortes 和 Vapnik 1995年; Breiman 2001年; Freund 和 Schapire 1997年) 不要只背算法清单,要比较它们的表示能力、优化目标、数据要求和失败模式。

强化学习以 Sutton 与 Barto 的教材为主线,再读 Q-learning 和 DQN 的论文。(Sutton 和 Barto 2018年; Watkins 和 Dayan 1992年; Mnih 等 2015年) 先在表格型环境中确认状态、动作、奖励、回报和探索,再进入函数逼近。若还不能解释“监督标签”和“延迟奖励”的差异,不宜直接从深度强化学习代码开始。

最小复现顺序:逻辑回归或决策树 -> 线性 SVM -> 小型集成模型 -> 网格世界 Q-learning -> 固定小游戏上的 DQN。每一步只增加一种主要复杂性。

路线三:深度学习与表示学习

配合第 13 章第 18 章,先用反向传播原始论文理解链式法则怎样在分层表示中复用,再阅读 LeNet、AlexNet、ReLU 初始化和 Adam。(Rumelhart 等 1986年; LeCun 等 1998年; Krizhevsky 等 2012年; He 等 2015年; Kingma 和 Ba 2015年)

这一组论文跨越近三十年,不能把性能变化只归因于“网络更深”。记录每项工作同时改变了哪些条件:数据规模、硬件、激活函数、正则化、优化器、工程实现和评测任务。模型架构只是一条因果链中的一环。

序列与生成路线可以从 LSTM、VAE 和 GAN 三篇代表论文进入。(Hochreiter 和 Schmidhuber 1997年; Kingma 和 Welling 2014年; Goodfellow 等 2014年) 比较三者时,不要只看样本图:写清楚似然或下界是否可计算、训练目标是否稳定、潜变量怎样使用,以及评测是否覆盖多样性和失真。

最小复现顺序:多层感知机与梯度检查 -> 小型 CNN -> LSTM 序列分类 -> 二维 VAE -> 小图像 GAN。Goodfellow 等的深度学习教材可作为跨主题参考,而不是必须顺序读完的大部头。(Goodfellow 等 2016年)

路线四:注意力、Transformer 与预训练

先读第 19 章第 21 章,再按问题演变阅读 Bahdanau 注意力、Transformer、BERT、GPT 和 T5。(Bahdanau 等 2015年; Vaswani 等 2017年; Devlin 等 2019年; Radford 等 2018年; Raffel 等 2020年)

第一次读 Transformer 论文,只追踪一个 Token 如何经过嵌入、位置表示、多头注意力、残差、归一化和前馈网络。第二次再核对矩阵形状、复杂度与掩码。第三次才比较编码器、解码器和编码器-解码器的训练目标。P08 的迷你 Transformer 实践适合验证这些局部机制。

预训练路线还需要补上规模规律与计算配比。GPT-3 展示上下文学习的规模效应,Kaplan 等和 Chinchilla 工作则给出不同条件下的经验缩放关系。(Brown 等 2020年; Kaplan 等 2020年; Hoffmann 等 2022年) 这些是特定模型族、数据和预算下的经验结果,不是无限外推的自然定律。基础模型综述适合把能力、同质化、数据、劳动、环境和治理风险放回同一张图。(Bommasani 等 2021年)

工程加餐:掌握标准注意力后再读 FlashAttention,区分数学函数不变与内存访问方式改变。(Dao 等 2022年)

路线五:生成模型、指令对齐与助手

扩散模型路线从第 22 章进入,依次阅读 DDPM、基于分数的生成建模和潜空间扩散。(Ho 等 2020年; Song 等 2021年; Rombach 等 2022年) 第一遍只回答前向过程加了什么噪声、反向过程预测什么;第二遍再比较参数化、采样步数、条件控制和计算位置。

助手路线从第 23 章进入。偏好学习、InstructGPT 和 DPO 分别提供了从人类比较中学习、构建指令助手和简化偏好优化的关键入口。(Christiano 等 2017年; Ouyang 等 2022年; Rafailov 等 2023年)

阅读对齐论文时,至少分开记录三件事:底层模型原有能力、训练后行为变化、部署系统额外限制。偏好数据代表某些标注者在某些任务上的选择,不自动等于事实正确、普遍价值或完整安全。

推荐实验:在小型分类或文本偏好数据上比较监督微调与成对偏好目标,观察奖励、准确性、输出长度和分布外表现是否一起变化。不要把一个代理指标的提升称为“完成对齐”。

路线六:多模态、推理、检索与智能体

第 25 章开始,用 CLIP 理解图文对比学习,再比较 Flamingo 与 BLIP-2 怎样连接视觉编码器和语言模型。(Radford 等 2021年; Alayrac 等 2022年; Li 等 2023年) 重点不是背组件名称,而是追踪视觉信息以什么表示进入语言生成,以及训练时哪些模块被冻结。

第 26 章对应的阅读顺序是思维链、自洽性、树搜索和过程验证。(Wei 等 2022年; Wang 等 2023年; Yao, Yu, 等 2023年; Lightman 等 2023年) 把“产生更多中间 Token”“搜索更多候选”和“拥有可靠推理能力”分开;测试时计算只有在候选具有多样性、评分器有辨别力且预算可接受时才可能转化为收益。

第 27 章第 28 章可沿 RAG、ReAct 与 Toolformer 阅读。(Lewis 等 2020年; Yao, Zhao, 等 2023年; Schick 等 2023年) 一次问答成功不能代表长任务可靠。SWE-bench 与 OSWorld 把评测推向真实代码仓库和计算机环境,也同时暴露环境配置、评分器、权限和可重复性问题。(Jimenez 等 2024年; Xie 等 2024年)

最小系统复现:固定文档集的检索问答 -> 带来源核对的回答 -> 两个只读工具 -> 有最大步数和失败恢复的循环。每次只开放一项能力,并记录端到端成功率,而不是只展示成功轨迹。

路线七:机器人与 AI for Science

第 29 章需要把模型输出重新放回感知、控制和安全闭环。RT-2 展示视觉语言模型与机器人动作的连接,DreamerV3 展示学习世界模型后进行决策的一条路线。(Brohan 等 2023年; Hafner 等 2023年) 阅读时检查训练环境、动作空间、控制频率、人工干预、失败恢复和现实测试次数;视频中的一次成功不足以给出可靠率。

第 30 章可按“预测 -> 搜索 -> 验证”组织来源。AlphaFold 2 与 AlphaFold 3 对应结构预测范围的扩展,GNoME 对应材料候选发现,GraphCast 对应天气预测,FunSearch 与 AlphaGeometry 对应程序和证明搜索。(Jumper 等 2021年; Abramson 等 2024年; Merchant 等 2023年; Lam 等 2023年; Romera-Paredes 等 2024年; Trinh 等 2024年)

最重要的阅读问题是:论文报告的是计算指标改善、候选被实验确认,还是已经改变科学与工程流程?模型提出候选不等于实验验证,实验验证也不等于规模化制造、临床有效或长期部署。

推荐产物:为一篇科学 AI 论文画证据链,明确数据来源、预测目标、搜索空间、验证手段和最终现实结论之间的每一步。

路线八:评测、安全、治理与社会影响

先读第 31 章,再把 MMLU、BIG-bench、HELM、GPQA 等材料当作“测量仪器说明书”阅读,而不是只抄分数。(Hendrycks 等 2021年; Srivastava 等 2023年; Liang 等 2023年; Rein 等 2024年) 代码评测可对读 HumanEval 与 EvalPlus,观察测试套件加严后结论怎样变化。(Chen 等 2021年; Liu 等 2023年)

第 32 章可结合模型卡、数据说明表、NIST AI 风险管理框架和通用语言模型风险分类。(Mitchell 等 2019年; Gebru 等 2021年; National Institute of Standards and Technology 2023年; Weidinger 等 2022年) 它们分别处理模型披露、数据治理、组织风险过程和风险空间,不能互相代替。

第 33 章涉及劳动和教育时,应区分任务暴露、采用率、生产率、就业数量、收入分配和工作质量。Autor 与 Acemoglu 提供自动化的长期分析框架,生成式 AI 的受控实验与现场研究则提供较近的任务级证据。(Autor 2015年; Acemoglu 和 Restrepo 2019年; Noy 和 Zhang 2023年; Brynjolfsson 等 2025年) 教育材料还应同时检查学习目标、评估方式、隐私和教师责任。(UNESCO 2023年)

推荐产物:选择一条现实新闻,用AI 新闻判断框架建立证据卡。将“技术可行”“组织采用”和“社会净影响”分别给出结论,不用一个模型分数代替三者。

路线九:AGI 与未来判断

第 34 章刻意不提供单一倒计时。Chollet 从技能获取效率讨论智能,Morris 等尝试用能力广度、表现等级和自主程度描述 AGI,Lake 等则从人类式学习能力提出研究缺口。(Chollet 2019年; Morris 等 2024年; Lake 等 2017年)

把这些定义与 Sutton 的“苦涩教训”、可扩展监督研究放在一起阅读,可以看到两类不同问题:怎样扩大一般方法的能力,以及能力扩大后怎样提供可扩展、可审计的人类约束。(Sutton 2019年; Leike 等 2018年)

阅读未来预测时,至少记录预测日期、目标定义、时间范围、前提、可观察指标和可能推翻它的事件。没有失败条件的预测更像立场表达,而不是可以更新的判断。

推荐产物:不要写一个年份,写一张状态表。分别评估任务广度、新任务适应、可靠性、现实锚定、资源效率和授权程度,并为每项注明证据日期。

怎样高效读一篇论文

第一遍:判断是否值得深读

先看标题、摘要、引言、主要图表、结论和局限。回答五个问题:

  1. 论文具体声称解决什么问题?
  2. 相对什么基线改善?
  3. 改变的是模型、数据、训练、推理预算还是评测?
  4. 最重要的证据在哪张表或哪幅图?
  5. 作者主动承认哪些边界?

若这五项还说不清,不要急着钻进推导。

第二遍:重建方法与比较条件

沿输入到输出画数据流,标出可学习参数、固定组件、损失和推理流程。再检查数据切分、基线实现、超参数搜索、运行次数、误差范围和消融实验。此时应能区分“方法本身带来的改善”和“更多数据、算力或重试带来的改善”。

第三遍:寻找会让结论变弱的证据

主动找失败案例、附录、负结果、误差条和未比较的强基线。随后搜索独立复现、后续批评和数据问题。目标不是驳倒论文,而是确定结论可以安全地走多远。

提示公式卡住时

先写出每个符号的类型和形状,再用一个两行三列的小例子手算。仍然不清楚时,回到公式前面的研究问题:这个量是在打分、归一化、累计证据,还是更新参数?数学附录提供了这套拆解方法。

来源笔记要能被未来的自己复查

每份来源建议保留一张简短卡片:

来源:作者、标题、年份、稳定标识符
版本:会议版 / 期刊版 / 预印本版本与日期
问题:它试图回答什么
对象:模型 / 系统 / 产品 / 部署 / 政策
方法:数据、结构、训练与推理条件
证据:最关键的表、图、实验或条款
结论:材料直接支持的最窄表述
边界:样本、分布、预算、失败模式与利益关系
复查:代码、数据、独立结果、勘误与访问日期
关联:本书章节及其他来源

直接引语要同时保存页码或段落位置;网页要记录标题、机构、发布日期和访问日期;会更新的排行榜要保存快照或精确版本。不要只保存一个可能失效的链接。

把自己的判断与来源原话分开写。一个实用标记法是:F 表示材料直接报告的事实,I 表示根据材料作出的推断,Q 表示尚未解决的问题。以后更新时,便能知道需要替换来源,还是需要修改自己的解释。

怎样核对引用与不断变化的事实

论文题名、作者、年份和标识符应以正式出版页、作者版本或可信索引为准。预印本后来进入会议或期刊时,记录两个版本之间是否改变了数据、实验和结论;不要仅改年份而默认内容相同。

模型、产品、排行榜、价格和政策最容易过期。引用它们时应同时保留:

  • 核查日期与地区;
  • 模型或产品的精确版本;
  • 基础模型还是带工具的系统;
  • 评测提示、采样、重试和预算;
  • 政策的法律状态、生效时间与适用对象;
  • 原页面修改或撤回后的替代来源。

至少用两个不同角色的来源核查强结论。例如,系统能力可以组合开发者报告与独立评测;社会影响可以组合任务实验与部署研究;政策可以组合正式文本与解释材料。两个转载同一新闻稿的网站仍然只有一个信息源。

逐章来源入口

下面给出的不是每章全部引用,而是适合从正文迈向原始材料的第一组入口。

每章末尾和全书参考文献仍保留更完整的来源。这里的入口只负责帮助读者迈出第一步。

一条十二周学习路线

这条路线按每周 5 至 8 小时设计,重点仍是理解内容,实践只占较小部分。时间不足时,可以把每一周延长为两周。

第 1 至 2 周:问题从哪里来

读序章和第 1 至 7 章;选读 Turing 1950、达特茅斯提案和一部 AI 史。产物是一页时间线和一页“规则系统的能力与边界”。

第 3 至 4 周:机器怎样从数据学习

读第 8 至 12 章和数学附录的概率、梯度部分;选读一本机器学习教材与强化学习教材的入门章节。可选实践是逻辑回归和网格世界 Q-learning。

第 5 至 6 周:深度网络怎样训练

读第 13 至 18 章;精读反向传播、CNN、LSTM、VAE 或 GAN 中两篇论文。产物是一张训练数据流和一份梯度检查记录。

第 7 至 8 周:从注意力到基础模型

读第 19 至 24 章;精读 Transformer,再在 BERT、GPT、扩散和对齐中选择两条支线。可选完成 P08,但只要求解释每个张量和掩码,不追求大规模训练。

第 9 至 10 周:模型怎样成为系统

读第 25 至 30 章;在多模态、推理、RAG、智能体、机器人和科学 AI 中选择两组原始材料。产物是一张包含工具、权限、验证和失败恢复的系统图。

第 11 周:怎样测量与治理

读第 31、32 章和一个基准的数据说明,再读模型卡、数据说明表或风险管理框架。选一张排行榜,重建它的任务、协议、评分和预算。

第 12 周:把结论放回社会与未来

读第 33、34 章以及本书三个判断型附录。选择一条当周 AI 新闻,制作来源卡和证据链;最后写下现在能确认什么、仍不知道什么、什么新证据会改变判断。

最后的阅读检查表

完成一个主题后,检查自己能否回答:

  • 我读的是历史记录、方法论文、系统报告、基准、标准,还是部署研究?
  • 核心主张的对象是模型、系统、产品、现实影响还是政策?
  • 输入、输出、训练目标和运行条件是什么?
  • 比较基线是否公平,数据是否可能污染,结果有没有不确定性?
  • 改善来自方法、数据、算力、工具、重试还是人工帮助?
  • 论文直接证明了什么,哪些只是作者或我自己的推断?
  • 结论在哪些分布、预算、版本和日期下成立?
  • 是否有独立复现、失败案例、后续修正或利益关系?
  • 我能否用一句更窄、更准确的话复述结论?
  • 下一份材料应该补机制、补反例,还是补现实部署证据?

延伸阅读的终点不是收藏更多链接,而是让自己的判断可以追溯、可以被反驳,也可以随着新证据更新。AI 会继续变化;一套知道该问什么、到哪里找证据、怎样缩小结论的方法,比任何静态的“最新清单”更耐用。

参考文献

Abramson, Josh, Jonas Adler, Jack Dunger, 等. 2024年. 《Accurate Structure Prediction of Biomolecular Interactions with AlphaFold 3. Nature 630: 493~500. https://doi.org/10.1038/s41586-024-07487-w.
Acemoglu, Daron, 和 Pascual Restrepo. 2019年. 《Automation and New Tasks: How Technology Displaces and Reinstates Labor》. Journal of Economic Perspectives 33 (2): 3~30. https://doi.org/10.1257/jep.33.2.3.
Alayrac, Jean-Baptiste, Jeff Donahue, Pauline Luc, 等. 2022年. 《Flamingo: A Visual Language Model for Few-Shot Learning》. Advances in Neural Information Processing Systems 35, 23716~36.
Autor, David H. 2015年. 《Why Are There Still So Many Jobs? The History and Future of Workplace Automation》. Journal of Economic Perspectives 29 (3): 3~30. https://doi.org/10.1257/jep.29.3.3.
Bahdanau, Dzmitry, Kyunghyun Cho, 和 Yoshua Bengio. 2015年. 《Neural Machine Translation by Jointly Learning to Align and Translate》. 3rd International Conference on Learning Representations. https://arxiv.org/abs/1409.0473.
Bishop, Christopher M. 2006年. Pattern Recognition and Machine Learning. Springer.
Bommasani, Rishi, Drew A. Hudson, Ehsan Adeli, 等. 2021年. 《On the Opportunities and Risks of Foundation Models》. arXiv preprint arXiv:2108.07258, 网络首发. https://doi.org/10.48550/arXiv.2108.07258.
Boole, George. 1854年. An Investigation of the Laws of Thought. Walton; Maberly. https://www.gutenberg.org/ebooks/15114.
Breiman, Leo. 2001年. 《Random Forests》. Machine Learning 45: 5~32. https://doi.org/10.1023/A:1010933404324.
Brohan, Anthony, Noah Brown, Justice Carbajal, 等. 2023年. RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control》. arXiv preprint arXiv:2307.15818.
Brown, Tom B., Benjamin Mann, Nick Ryder, 等. 2020年. 《Language Models are Few-Shot Learners》. Advances in Neural Information Processing Systems 33, 1877~901.
Brynjolfsson, Erik, Danielle Li, 和 Lindsey R. Raymond. 2025年. 《Generative AI at Work》. The Quarterly Journal of Economics 140 (2): 889~942. https://doi.org/10.1093/qje/qjae044.
Chen, Mark, Jerry Tworek, Heewoo Jun, 等. 2021年. 《Evaluating Large Language Models Trained on Code》. arXiv preprint arXiv:2107.03374.
Chollet, Francois. 2019年. 《On the Measure of Intelligence》. arXiv preprint arXiv:1911.01547.
Christiano, Paul F., Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, 和 Dario Amodei. 2017年. 《Deep Reinforcement Learning from Human Preferences》. Advances in Neural Information Processing Systems 30.
Church, Alonzo. 1936年. 《An Unsolvable Problem of Elementary Number Theory》. American Journal of Mathematics 58 (2): 345~63. https://doi.org/10.2307/2371045.
Cortes, Corinna, 和 Vladimir Vapnik. 1995年. 《Support-Vector Networks》. Machine Learning 20: 273~97. https://doi.org/10.1007/BF00994018.
Dao, Tri, Daniel Y. Fu, Stefano Ermon, Atri Rudra, 和 Christopher Ré. 2022年. 《FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness》. Advances in Neural Information Processing Systems 35, 16344~59. https://papers.nips.cc/paper_files/paper/2022/hash/67d57c32e20fd0a7a302cb81d36e40d5-Abstract-Conference.html.
Devlin, Jacob, Ming-Wei Chang, Kenton Lee, 和 Kristina Toutanova. 2019年. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding》. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics, 4171~86. https://doi.org/10.18653/v1/N19-1423.
Freund, Yoav, 和 Robert E. Schapire. 1997年. 《A Decision-Theoretic Generalization of On-Line Learning and an Application to Boosting》. Journal of Computer and System Sciences 55 (1): 119~39. https://doi.org/10.1006/jcss.1997.1504.
Gebru, Timnit, Jamie Morgenstern, Briana Vecchione, 等. 2021年. 《Datasheets for Datasets》. Communications of the ACM 64 (12): 86~92. https://doi.org/10.1145/3458723.
Goodfellow, Ian, Yoshua Bengio, 和 Aaron Courville. 2016年. Deep Learning. MIT Press. https://www.deeplearningbook.org/.
Goodfellow, Ian, Jean Pouget-Abadie, Mehdi Mirza, 等. 2014年. 《Generative Adversarial Nets》. Advances in Neural Information Processing Systems 27, 2672~80.
Hafner, Danijar, Jurgis Pasukonis, Jimmy Ba, 和 Timothy Lillicrap. 2023年. 《Mastering Diverse Domains through World Models》. arXiv preprint arXiv:2301.04104.
Hastie, Trevor, Robert Tibshirani, 和 Jerome Friedman. 2009年. The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 2 版. Springer. https://doi.org/10.1007/978-0-387-84858-7.
He, Kaiming, Xiangyu Zhang, Shaoqing Ren, 和 Jian Sun. 2015年. 《Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification》. Proceedings of the IEEE International Conference on Computer Vision, 1026~34. https://doi.org/10.1109/ICCV.2015.123.
Hendrycks, Dan, Collin Burns, Steven Basart, 等. 2021年. 《Measuring Massive Multitask Language Understanding》. 9th International Conference on Learning Representations. https://arxiv.org/abs/2009.03300.
Ho, Jonathan, Ajay Jain, 和 Pieter Abbeel. 2020年. 《Denoising Diffusion Probabilistic Models》. Advances in Neural Information Processing Systems 33, 6840~51.
Hochreiter, Sepp, 和 Jürgen Schmidhuber. 1997年. 《Long Short-Term Memory》. Neural Computation 9 (8): 1735~80. https://doi.org/10.1162/neco.1997.9.8.1735.
Hoffmann, Jordan, Sebastian Borgeaud, Arthur Mensch, 等. 2022年. 《Training Compute-Optimal Large Language Models》. Advances in Neural Information Processing Systems 35, 30016~30.
Jimenez, Carlos E., John Yang, Alexander Wettig, 等. 2024年. SWE-bench: Can Language Models Resolve Real-World GitHub Issues?》 12th International Conference on Learning Representations. https://arxiv.org/abs/2310.06770.
Jumper, John, Richard Evans, Alexander Pritzel, 等. 2021年. 《Highly Accurate Protein Structure Prediction with AlphaFold. Nature 596: 583~89. https://doi.org/10.1038/s41586-021-03819-2.
Kaplan, Jared, Sam McCandlish, Tom Henighan, 等. 2020年. 《Scaling Laws for Neural Language Models》. arXiv preprint arXiv:2001.08361, 网络首发. https://doi.org/10.48550/arXiv.2001.08361.
Kingma, Diederik P., 和 Jimmy Ba. 2015年. 《Adam: A Method for Stochastic Optimization》. 3rd International Conference on Learning Representations. https://arxiv.org/abs/1412.6980.
Kingma, Diederik P., 和 Max Welling. 2014年. 《Auto-Encoding Variational Bayes》. 2nd International Conference on Learning Representations. https://arxiv.org/abs/1312.6114.
Krizhevsky, Alex, Ilya Sutskever, 和 Geoffrey E. Hinton. 2012年. ImageNet Classification with Deep Convolutional Neural Networks》. Advances in Neural Information Processing Systems 25, 1097~105.
Lake, Brenden M., Tomer D. Ullman, Joshua B. Tenenbaum, 和 Samuel J. Gershman. 2017年. 《Building Machines That Learn and Think Like People》. Behavioral and Brain Sciences 40: e253. https://doi.org/10.1017/S0140525X16001837.
Lam, Remi, Alvaro Sanchez-Gonzalez, Matthew Willson, 等. 2023年. 《Learning Skillful Medium-Range Global Weather Forecasting》. Science 382 (6677): 1416~21. https://doi.org/10.1126/science.adi2336.
LeCun, Yann, Leon Bottou, Yoshua Bengio, 和 Patrick Haffner. 1998年. 《Gradient-Based Learning Applied to Document Recognition》. Proceedings of the IEEE 86 (11): 2278~324. https://doi.org/10.1109/5.726791.
Leike, Jan, David Krueger, Tom Everitt, Miljan Martic, Vishal Maini, 和 Shane Legg. 2018年. 《Scalable Agent Alignment via Reward Modeling: A Research Direction》. arXiv preprint arXiv:1811.07871.
Lewis, Patrick, Ethan Perez, Aleksandra Piktus, 等. 2020年. 《Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks》. Advances in Neural Information Processing Systems 33. https://arxiv.org/abs/2005.11401.
Li, Junnan, Dongxu Li, Silvio Savarese, 和 Steven Hoi. 2023年. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models》. Proceedings of the 40th International Conference on Machine Learning, Proceedings of Machine Learning Research, 卷 202: 19730~42.
Liang, Percy, Rishi Bommasani, Tony Lee, 等. 2023年. 《Holistic Evaluation of Language Models》. Transactions on Machine Learning Research. https://arxiv.org/abs/2211.09110.
Lightman, Hunter, Vineet Kosaraju, Yura Burda, 等. 2023年. 《Let’s Verify Step by Step》. arXiv preprint arXiv:2305.20050, 网络首发. https://doi.org/10.48550/arXiv.2305.20050.
Liu, Jiawei, Chunqiu Steven Xia, Yuyao Wang, 和 Lingming Zhang. 2023年. 《Is Your Code Generated by ChatGPT Really Correct? Rigorous Evaluation of Large Language Models for Code Generation》. Advances in Neural Information Processing Systems 36. https://arxiv.org/abs/2305.01210.
Lovelace, Augusta Ada. 1843年. 《Sketch of the Analytical Engine Invented by Charles Babbage, with Notes by the Translator》. Scientific Memoirs 3: 666~731. https://www.fourmilab.ch/babbage/sketch.html.
McCarthy, John, Marvin L. Minsky, Nathaniel Rochester, 和 Claude E. Shannon. 1955年. A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. http://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html.
Merchant, Amil, Simon Batzner, Samuel S. Schoenholz, Muratahan Aykol, Gowoon Cheon, 和 Ekin Dogus Cubuk. 2023年. 《Scaling Deep Learning for Materials Discovery》. Nature 624: 80~85. https://doi.org/10.1038/s41586-023-06735-9.
Mitchell, Margaret, Simone Wu, Andrew Zaldivar, 等. 2019年. 《Model Cards for Model Reporting》. Proceedings of the Conference on Fairness, Accountability, and Transparency, 220~29. https://doi.org/10.1145/3287560.3287596.
Mitchell, Tom M. 1997年. Machine Learning. McGraw-Hill.
Mnih, Volodymyr, Koray Kavukcuoglu, David Silver, 等. 2015年. 《Human-Level Control through Deep Reinforcement Learning》. Nature 518: 529~33. https://doi.org/10.1038/nature14236.
Morris, Meredith Ringel, Jascha Sohl-Dickstein, Noah Fiedel, 等. 2024年. 《Position: Levels of AGI for Operationalizing Progress on the Path to AGI. Proceedings of the 41st International Conference on Machine Learning. https://arxiv.org/abs/2311.02462.
National Institute of Standards and Technology. 2023年. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. National Institute of Standards; Technology. https://doi.org/10.6028/NIST.AI.100-1.
Newell, Allen, 和 Herbert A. Simon. 1956年. 《The Logic Theory Machine: A Complex Information Processing System》. IRE Transactions on Information Theory 2 (3): 61~79. https://doi.org/10.1109/TIT.1956.1056797.
Nilsson, Nils J. 2010年. The Quest for Artificial Intelligence: A History of Ideas and Achievements. Cambridge University Press. https://doi.org/10.1017/CBO9780511819346.
Noy, Shakked, 和 Whitney Zhang. 2023年. 《Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence》. Science 381 (6654): 187~92. https://doi.org/10.1126/science.adh2586.
Ouyang, Long, Jeffrey Wu, Xu Jiang, 等. 2022年. 《Training Language Models to Follow Instructions with Human Feedback》. Advances in Neural Information Processing Systems 35, 27730~44.
Radford, Alec, Jong Wook Kim, Chris Hallacy, 等. 2021年. 《Learning Transferable Visual Models From Natural Language Supervision》. Proceedings of the 38th International Conference on Machine Learning, Proceedings of Machine Learning Research, 卷 139: 8748~63.
Radford, Alec, Karthik Narasimhan, Tim Salimans, 和 Ilya Sutskever. 2018年. Improving Language Understanding by Generative Pre-Training. OpenAI. https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
Rafailov, Rafael, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, 和 Chelsea Finn. 2023年. 《Direct Preference Optimization: Your Language Model is Secretly a Reward Model》. Advances in Neural Information Processing Systems 36.
Raffel, Colin, Noam Shazeer, Adam Roberts, 等. 2020年. 《Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer》. Journal of Machine Learning Research 21 (140): 1~67. https://jmlr.org/papers/v21/20-074.html.
Rein, David, Betty Li Hou, Asa Cooper Stickland, 等. 2024年. GPQA: A Graduate-Level Google-Proof Q&A Benchmark》. Proceedings of the First Conference on Language Modeling. https://arxiv.org/abs/2311.12022.
Rombach, Robin, Andreas Blattmann, Dominik Lorenz, Patrick Esser, 和 Bjorn Ommer. 2022年. 《High-Resolution Image Synthesis with Latent Diffusion Models》. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 10684~95. https://doi.org/10.1109/CVPR52688.2022.01042.
Romera-Paredes, Bernardino, Mohammadamin Barekatain, Alexander Novikov, 等. 2024年. 《Mathematical Discoveries from Program Search with Large Language Models》. Nature 625: 468~75. https://doi.org/10.1038/s41586-023-06924-6.
Rosenblatt, Frank. 1958年. 《The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain》. Psychological Review 65 (6): 386~408. https://doi.org/10.1037/h0042519.
Rumelhart, David E., Geoffrey E. Hinton, 和 Ronald J. Williams. 1986年. 《Learning Representations by Back-Propagating Errors》. Nature 323: 533~36. https://doi.org/10.1038/323533a0.
Schick, Timo, Jane Dwivedi-Yu, Roberto Dessi, 等. 2023年. Toolformer: Language Models Can Teach Themselves to Use Tools》. Advances in Neural Information Processing Systems 36. https://arxiv.org/abs/2302.04761.
Shannon, Claude E. 1938年. 《A Symbolic Analysis of Relay and Switching Circuits》. Transactions of the American Institute of Electrical Engineers 57 (12): 713~23. https://doi.org/10.1109/T-AIEE.1938.5057767.
Shannon, Claude E. 1948年. 《A Mathematical Theory of Communication》. Bell System Technical Journal 27 (3–4): 379~423, 623~56. https://doi.org/10.1002/j.1538-7305.1948.tb01338.x.
Song, Yang, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, 和 Ben Poole. 2021年. 《Score-Based Generative Modeling through Stochastic Differential Equations》. 9th International Conference on Learning Representations.
Srivastava, Aarohi, Abhinav Rastogi, Abhishek Rao, 等. 2023年. 《Beyond the Imitation Game: Quantifying and Extrapolating the Capabilities of Language Models》. Transactions on Machine Learning Research. https://arxiv.org/abs/2206.04615.
Sutton, Richard S. 2019年. The Bitter Lesson. http://www.incompleteideas.net/IncIdeas/BitterLesson.html.
Sutton, Richard S., 和 Andrew G. Barto. 2018年. Reinforcement Learning: An Introduction. 2 版. MIT Press. http://incompleteideas.net/book/the-book-2nd.html.
Trinh, Trieu H., Yuhuai Wu, Quoc V. Le, He He, 和 Thang Luong. 2024年. 《Solving Olympiad Geometry without Human Demonstrations》. Nature 625: 476~82. https://doi.org/10.1038/s41586-023-06747-5.
Turing, Alan M. 1937年. 《On Computable Numbers, with an Application to the Entscheidungsproblem》. Proceedings of the London Mathematical Society s2-42 (1): 230~65. https://doi.org/10.1112/plms/s2-42.1.230.
Turing, Alan M. 1950年. 《Computing Machinery and Intelligence》. Mind LIX (236): 433~60. https://doi.org/10.1093/mind/LIX.236.433.
UNESCO. 2023年. Guidance for Generative AI in Education and Research. UNESCO.
Vapnik, Vladimir N. 1995年. The Nature of Statistical Learning Theory. Springer. https://doi.org/10.1007/978-1-4757-2440-0.
Vaswani, Ashish, Noam Shazeer, Niki Parmar, 等. 2017年. 《Attention Is All You Need》. Advances in Neural Information Processing Systems 30, 5998~6008. https://papers.nips.cc/paper/7181-attention-is-all-you-need.
Wang, Xuezhi, Jason Wei, Dale Schuurmans, 等. 2023年. 《Self-Consistency Improves Chain of Thought Reasoning in Language Models》. 11th International Conference on Learning Representations.
Watkins, Christopher J. C. H., 和 Peter Dayan. 1992年. 《Q-Learning》. Machine Learning 8: 279~92. https://doi.org/10.1007/BF00992698.
Wei, Jason, Xuezhi Wang, Dale Schuurmans, 等. 2022年. 《Chain-of-Thought Prompting Elicits Reasoning in Large Language Models》. Advances in Neural Information Processing Systems 35, 24824~37.
Weidinger, Laura, John Mellor, Maribeth Rauh, 等. 2022年. 《Taxonomy of Risks Posed by Language Models》. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, 214~29. https://doi.org/10.1145/3531146.3533088.
Wiener, Norbert. 1948年. Cybernetics: Or Control and Communication in the Animal and the Machine. MIT Press.
Xie, Tianbao, Danyang Zhang, Jixuan Chen, 等. 2024年. OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments》. Advances in Neural Information Processing Systems 37. https://arxiv.org/abs/2404.07972.
Yao, Shunyu, Dian Yu, Jeffrey Zhao, 等. 2023年. 《Tree of Thoughts: Deliberate Problem Solving with Large Language Models》. Advances in Neural Information Processing Systems 36.
Yao, Shunyu, Jeffrey Zhao, Dian Yu, 等. 2023年. ReAct: Synergizing Reasoning and Acting in Language Models》. 11th International Conference on Learning Representations. https://arxiv.org/abs/2210.03629.