[
  {
    "case": "P1",
    "slug": "instructgpt",
    "title": "InstructGPT",
    "arxiv_id": "2203.02155",
    "figure": "2",
    "pdf_page": 3,
    "source_url": "https://arxiv.org/pdf/2203.02155",
    "source_pdf_sha256": "c1984bb50a5b90fddb895fdc3a0f72e5bc977148c9f63ef6040cbe7a3e1f0d98",
    "crop_pdf_points": [
      106,
      63,
      506,
      300
    ],
    "dpi": 300,
    "width": 1668,
    "height": 988,
    "caption": "Figure 2: A diagram illustrating the three steps of our method: (1) supervised ﬁne-tuning (SFT), (2)\nreward model (RM) training, and (3) reinforcement learning via proximal policy optimization (PPO)\non this reward model. Blue arrows indicate that this data is used to train one of our models. In Step 2,\nboxes A-D are samples from our models that get ranked by labelers. See Section 3 for more details\non our method.",
    "comparison_note": "原图三阶段横排，包含标注员、模型训练和采样；重绘仅保留阶段骨架。",
    "png": "instructgpt.png",
    "pdf": "instructgpt.pdf",
    "redraw": "instructgpt-redraw.svg",
    "method": "PDF region rendering; original text, vectors and images retained; no redrawing",
    "rights": "Original figure copyright belongs to the paper rights holders; site code license does not apply."
  },
  {
    "case": "P2",
    "slug": "bert",
    "title": "BERT",
    "arxiv_id": "1810.04805",
    "figure": "1",
    "pdf_page": 3,
    "source_url": "https://aclanthology.org/N19-1423.pdf",
    "source_pdf_sha256": "987545ffb087f1ece898142c403a516baeabeb70ce19089397fac6f7db12c3d4",
    "crop_pdf_points": [
      72,
      66,
      526,
      250
    ],
    "dpi": 300,
    "width": 1892,
    "height": 767,
    "caption": "Figure 1: Overall pre-training and ﬁne-tuning procedures for BERT. Apart from output layers, the same architec-\ntures are used in both pre-training and ﬁne-tuning. The same pre-trained model parameters are used to initialize\nmodels for different down-stream tasks. During ﬁne-tuning, all parameters are ﬁne-tuned. [CLS] is a special\nsymbol added in front of every input example, and [SEP] is a special separator token (e.g. separating ques-\ntions/answers).",
    "comparison_note": "原图含预训练任务头和多个微调任务；重绘只保留双阶段关系。",
    "png": "bert.png",
    "pdf": "bert.pdf",
    "redraw": "bert-redraw.svg",
    "method": "PDF region rendering; original text, vectors and images retained; no redrawing",
    "rights": "Original figure copyright belongs to the paper rights holders; site code license does not apply."
  },
  {
    "case": "P3",
    "slug": "transformer",
    "title": "Transformer",
    "arxiv_id": "1706.03762",
    "figure": "1",
    "pdf_page": 3,
    "source_url": "https://arxiv.org/pdf/1706.03762",
    "source_pdf_sha256": "bdfaa68d8984f0dc02beaca527b76f207d99b666d31d1da728ee0728182df697",
    "crop_pdf_points": [
      190,
      62,
      423,
      399
    ],
    "dpi": 300,
    "width": 972,
    "height": 1405,
    "caption": "Figure 1: The Transformer - model architecture.",
    "comparison_note": "原图保留完整编解码子层、残差与归一化；重绘压缩了这些结构。",
    "png": "transformer.png",
    "pdf": "transformer.pdf",
    "redraw": "transformer-redraw.svg",
    "method": "PDF region rendering; original text, vectors and images retained; no redrawing",
    "rights": "Original figure copyright belongs to the paper rights holders; site code license does not apply."
  },
  {
    "case": "P4",
    "slug": "clip",
    "title": "CLIP",
    "arxiv_id": "2103.00020",
    "figure": "1",
    "pdf_page": 2,
    "source_url": "https://proceedings.mlr.press/v139/radford21a/radford21a.pdf",
    "source_pdf_sha256": "2a4952960f350e50a3a4a847371b80b8aaa09ba82d1c2a8f6b30450e392a1fe4",
    "crop_pdf_points": [
      53,
      64,
      545,
      245
    ],
    "dpi": 300,
    "width": 2051,
    "height": 755,
    "caption": "Figure 1. Summary of our approach. While standard image models jointly train an image feature extractor and a linear classiﬁer to predict",
    "comparison_note": "原图还含文本构造分类器和零样本预测两步；重绘主要展示对比训练。",
    "png": "clip.png",
    "pdf": "clip.pdf",
    "redraw": "clip-redraw.svg",
    "method": "PDF region rendering; original text, vectors and images retained; no redrawing",
    "rights": "Original figure copyright belongs to the paper rights holders; site code license does not apply."
  },
  {
    "case": "P5",
    "slug": "ddpm",
    "title": "DDPM",
    "arxiv_id": "2006.11239",
    "figure": "2",
    "pdf_page": 2,
    "source_url": "https://arxiv.org/pdf/2006.11239",
    "source_pdf_sha256": "aee5e07a802e8dfd2a386374c94fd61d1d056cb7e1e0fec4f28e8120ff5d8505",
    "crop_pdf_points": [
      158,
      68,
      455,
      122
    ],
    "dpi": 300,
    "width": 1238,
    "height": 226,
    "caption": "Figure 2: The directed graphical model considered in this work.",
    "comparison_note": "原图 xT 在左、x0 在右；生成方向向右、加噪方向向左。重绘改变了状态排列，不能照搬其方向说明。",
    "png": "ddpm.png",
    "pdf": "ddpm.pdf",
    "redraw": "ddpm-redraw.svg",
    "method": "PDF region rendering; original text, vectors and images retained; no redrawing",
    "rights": "Original figure copyright belongs to the paper rights holders; site code license does not apply."
  },
  {
    "case": "P6",
    "slug": "vit",
    "title": "Vision Transformer",
    "arxiv_id": "2010.11929",
    "figure": "1",
    "pdf_page": 3,
    "source_url": "https://arxiv.org/pdf/2010.11929",
    "source_pdf_sha256": "8ce7b83971a14508ca711a27c875c9b6914c4f6767cf3150fb1ca6c07aa056d6",
    "crop_pdf_points": [
      130,
      75,
      480,
      256
    ],
    "dpi": 300,
    "width": 1459,
    "height": 755,
    "caption": "Figure 1: Model overview. We split an image into ﬁxed-size patches, linearly embed each of them,\nadd position embeddings, and feed the resulting sequence of vectors to a standard Transformer\nencoder. In order to perform classiﬁcation, we use the standard approach of adding an extra learnable\n“classiﬁcation token” to the sequence. The illustration of the Transformer encoder was inspired by\nVaswani et al. (2017).",
    "comparison_note": "原图同时展开 Transformer Encoder 内部与分类头；重绘强调 patch 转 token。",
    "png": "vit.png",
    "pdf": "vit.pdf",
    "redraw": "vit-redraw.svg",
    "method": "PDF region rendering; original text, vectors and images retained; no redrawing",
    "rights": "Original figure copyright belongs to the paper rights holders; site code license does not apply."
  },
  {
    "case": "P7",
    "slug": "gpt3",
    "title": "GPT-3",
    "arxiv_id": "2005.14165",
    "figure": "3.8",
    "pdf_page": 20,
    "source_url": "https://arxiv.org/pdf/2005.14165",
    "source_pdf_sha256": "97fd272f1fdfc18677462d0292f5fbf26ca86b4d1b485c2dba03269b643a0e83",
    "crop_pdf_points": [
      71,
      65,
      543,
      290
    ],
    "dpi": 300,
    "width": 1968,
    "height": 939,
    "caption": "Figure 3.8: Performance on SuperGLUE increases with model size and number of examples in context. A value\nof K = 32 means that our model was shown 32 examples per task, for 256 examples total divided across the 8 tasks in\nSuperGLUE. We report GPT-3 values on the dev set, so our numbers are not directly comparable to the dotted reference\nlines (our test set results are in Table 3.8). The BERT-Large reference model was ﬁne-tuned on the SuperGLUE training\nset (125K examples), whereas BERT++ was ﬁrst ﬁne-tuned on MultiNLI (392K examples) and SWAG (113K examples)\nbefore further ﬁne-tuning on the SuperGLUE training set (for a total of 630K ﬁne-tuning examples). We ﬁnd the\ndifference in performance between the BERT-Large and BERT++ to be roughly equivalent to the difference between\nGPT-3 with one example per context versus eight examples per context.",
    "comparison_note": "选用 Figure 3.8：左图是模型规模，右图是上下文示例数。原图是两个面板；重绘多任务曲线并非逐点复现，示意数值不可当作原论文结果。",
    "png": "gpt3.png",
    "pdf": "gpt3.pdf",
    "redraw": "gpt3-redraw.svg",
    "method": "PDF region rendering; original text, vectors and images retained; no redrawing",
    "rights": "Original figure copyright belongs to the paper rights holders; site code license does not apply."
  },
  {
    "case": "P8",
    "slug": "dpo",
    "title": "DPO",
    "arxiv_id": "2305.18290",
    "figure": "1",
    "pdf_page": 2,
    "source_url": "https://arxiv.org/pdf/2305.18290",
    "source_pdf_sha256": "92cb3a2b71362acda98a789b03d88688fd33cf5fcf13f81d2b1de30ee7d3b67a",
    "crop_pdf_points": [
      106,
      69,
      507,
      154
    ],
    "dpi": 300,
    "width": 1672,
    "height": 355,
    "caption": "Figure 1: DPO optimizes for human preferences while avoiding reinforcement learning. Existing methods\nfor fine-tuning language models with human feedback first fit a reward model to a dataset of prompts and\nhuman preferences over pairs of responses, and then use RL to find a policy that maximizes the learned reward.\nIn contrast, DPO directly optimizes for the policy best satisfying the preferences with a simple classification\nobjective, fitting an implicit reward model whose corresponding optimal policy can be extracted in closed form.",
    "comparison_note": "原图 RLHF 与 DPO 左右并排；重绘改为上下两行。原图包含采样与奖励计算回路。",
    "png": "dpo.png",
    "pdf": "dpo.pdf",
    "redraw": "dpo-redraw.svg",
    "method": "PDF region rendering; original text, vectors and images retained; no redrawing",
    "rights": "Original figure copyright belongs to the paper rights holders; site code license does not apply."
  },
  {
    "case": "P9",
    "slug": "switch",
    "title": "Switch Transformer",
    "arxiv_id": "2101.03961",
    "figure": "2",
    "pdf_page": 4,
    "source_url": "https://www.jmlr.org/papers/volume23/21-0998/21-0998.pdf",
    "source_pdf_sha256": "8831edca732dd939854957b4b1a493fa553e7b43dd431a5d4fee15493bba6e3e",
    "crop_pdf_points": [
      90,
      86,
      523,
      297
    ],
    "dpi": 300,
    "width": 1805,
    "height": 880,
    "caption": "Figure 2: Illustration of a Switch Transformer encoder block. We replace the dense feed",
    "comparison_note": "原图左侧是编码器模块定位，右侧展开稀疏 FFN；左侧不是密集 FFN 对照。每个 token 选一个专家，两个 token 可选择不同专家。",
    "png": "switch.png",
    "pdf": "switch.pdf",
    "redraw": "switch-redraw.svg",
    "method": "PDF region rendering; original text, vectors and images retained; no redrawing",
    "rights": "Original figure copyright belongs to the paper rights holders; site code license does not apply."
  },
  {
    "case": "P10",
    "slug": "cot",
    "title": "Chain-of-Thought",
    "arxiv_id": "2201.11903",
    "figure": "1",
    "pdf_page": 1,
    "source_url": "https://arxiv.org/pdf/2201.11903",
    "source_pdf_sha256": "7d9f878c23b460e4566aa4ec9201b1abfb3b8faefb2b1356e411cb90fef72a12",
    "crop_pdf_points": [
      115,
      450,
      501,
      642
    ],
    "dpi": 300,
    "width": 1609,
    "height": 800,
    "caption": "Figure 1: Chain-of-thought prompting enables large language models to tackle complex arithmetic,\ncommonsense, and symbolic reasoning tasks. Chain-of-thought reasoning processes are highlighted.",
    "comparison_note": "原图包含网球示例和食堂测试题两组输入输出；重绘只保留提示对照。原图最终答案为 27 与 9。",
    "png": "cot.png",
    "pdf": "cot.pdf",
    "redraw": "cot-redraw.svg",
    "method": "PDF region rendering; original text, vectors and images retained; no redrawing",
    "rights": "Original figure copyright belongs to the paper rights holders; site code license does not apply."
  }
]
