Back to Details

ljg-paper-river

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

ljg-paper-connects: 倒读法

ljg-paper-connects: Reverse Reading Method

一篇论文不是孤岛。它站在前人的肩上，也踩着前人的伤疤。倒着挖到根，再正着看过来——问题怎么长出来的，每个人看到了什么别人没看到的，解法怎么一步步逼近真相。
A research paper is not an isolated island. It stands on the shoulders of predecessors, and also steps on the flaws of previous work. Dig backwards to the root, then look forward from there - see how the problem emerged, what each researcher noticed that others missed, and how solutions gradually approached the truth.

核心逻辑

Core Logic

读论文最常见的错：只看眼前这一篇，不知道它从哪来。倒读法反过来——先找到这篇论文在批判谁、改进谁，再找那篇论文又在批判谁，递归五层，挖到源头。然后掉头，从源头正向读回来。
这样读完，你拿到的不是一篇论文的知识，是一整条问题演化线的理解。
The most common mistake when reading papers: only reading the current paper without knowing its background. The reverse reading method works the other way around - first find which papers this paper critiques and improves on, then find which papers those papers critique and improve on, recursively for 5 layers until you reach the origin. Then turn around and read forward from the origin.
After reading this way, you won't just get the knowledge of a single paper, but a complete understanding of the entire evolution line of the problem.

格式约束

Format Constraints

Org-mode 语法

Org-mode Syntax

  • 加粗用 
    *bold*
    （单星号），禁止 
    **bold**
  • 标题层级从 
    *
    开始，不跳级
  • Bold uses 
    *bold*
    (single asterisk), 
    **bold**
    is prohibited
  • Heading levels start from 
    *
    , no level skipping

ASCII Art

ASCII Art

所有图表用纯 ASCII 字符。允许：
+ - | / \ > < v ^ * = ~ . : # [ ] ( ) _ , ; ! ' "
和空格。禁止 Unicode 绘图符号。
All charts use pure ASCII characters. Allowed characters: 
+ - | / \ > < v ^ * = ~ . : # [ ] ( ) _ , ; ! ' "
and spaces. Unicode drawing symbols are prohibited.

模板权威性

Template Authority

输出结构依据 
references/template.org
。
Output structure follows 
references/template.org
.

Denote 文件规范

Denote File Specifications

  • 时间戳：
    date +%Y%m%dT%H%M%S
  • 可读时间：
    date "+%Y-%m-%d %a %H:%M"
  • 文件名：
    {时间戳}--paper-river-{简短标题}__paper_river.org
  • 输出目录：
    ~/Documents/notes/
  • Timestamp: 
    date +%Y%m%dT%H%M%S
  • Readable time: 
    date "+%Y-%m-%d %a %H:%M"
  • File name: 
    {timestamp}--paper-river-{short-title}__paper_river.org
  • Output directory: 
    ~/Documents/notes/

Org 文件头

Org File Header

#+title:      paper-river-{简短标题}
#+date:       [{YYYY-MM-DD Day HH:MM}]
#+filetags:   :paper:river:
#+identifier: {YYYYMMDDTHHMMSS}
#+source:     {URL 或来源描述}
#+authors:    {目标论文作者}
#+venue:      {发表场所/年份}
#+title:      paper-river-{short-title}
#+date:       [{YYYY-MM-DD Day HH:MM}]
#+filetags:   :paper:river:
#+identifier: {YYYYMMDDTHHMMSS}
#+source:     {URL or source description}
#+authors:    {target paper authors}
#+venue:      {publication venue/year}

红线

Red Lines

  1. 问题为轴 — 整篇文章的主线是"问题怎么演化的"，不是"论文怎么排列的"。论文是配角，问题是主角
  2. 口语检验 — 你会这样跟朋友讲一个领域的发展史吗？不会就改
  3. 差异为核 — 每篇论文的讲解重心是"它和前一篇的差异在哪"，不是独立地介绍每篇论文
  4. 零术语 — 先用大白话落地，再顺带提术语名
  5. 逻辑不断链 — 从第一篇到最后一篇，因果链条不能断。读者能感受到"所以他们才会这样做"
  6. 诚实 — 找不到五层就说找到几层。论文之间的关系不确定就说不确定。不编造引用关系
  1. Problem-centric — The main line of the entire article is "how the problem evolved", not "how papers are arranged". Papers are supporting roles, problems are the main character
  2. Oral language test — Would you explain the development history of a field to a friend like this? Revise if not
  3. Difference as core — The focus of each paper's explanation is "what is the difference between it and the previous one", not introducing each paper independently
  4. Zero jargon first — Explain in plain language first, then mention the terminology incidentally
  5. Unbroken logical chain — From the first paper to the last, the causal chain cannot be broken. Readers should feel "that's why they did it this way"
  6. Honesty — State how many layers you found if you can't find 5 layers. State if the relationship between papers is uncertain. Do not fabricate citation relationships

写作原则

Writing Principles

  1. 差异驱动叙事 — 不要给每篇论文写独立摘要再拼起来。以"这篇看到了前一篇的什么问题"作为每段的开头，让差异本身推动叙事往前走
  2. 变形替代定义 — 讲两个方案的区别时，把方案A连续变形成方案B。"如果你把X去掉，再加上Y，你就得到了Z"——比"Z和X的区别是..."有力十倍
  3. 推理外显 — 每个解法出现前，先让读者感受到"不这么做不行了"的压力。模拟发现的过程，不是汇报发现的结果
  4. 一张图胜千言 — 在演化叙事之前画溯源地图，在叙事之后画压缩总览图。让读者先有全景再入细节，细节看完再回全景
  1. Difference-driven narrative — Do not write independent abstracts for each paper and piece them together. Start each paragraph with "what problem did this paper find in the previous one", let the differences themselves drive the narrative forward
  2. Transformation instead of definition — When explaining the difference between two solutions, continuously transform solution A into solution B. "If you remove X, add Y, you get Z" — ten times more powerful than "the difference between Z and X is..."
  3. Explicit reasoning — Before each solution appears, let readers feel the pressure that "we have to do it this way". Simulate the discovery process, not report the discovery result
  4. A picture is worth a thousand words — Draw a traceability map before the evolution narrative, and draw a compressed overview map after the narrative. Let readers have a panoramic view first before entering details, and return to the panoramic view after reading the details

执行

Execution

1. 获取目标论文

1. Obtain the target paper

  • arxiv URL → WebFetch
  • PDF → Read（注意 pages 参数限制）
  • 论文名称 → WebSearch 找到全文
确保拿到：标题、作者、摘要、引言（尤其是 related work / introduction 中对前人工作的批判）。
  • arxiv URL → WebFetch
  • PDF → Read (pay attention to pages parameter limit)
  • Paper name → WebSearch to find the full text
Make sure to get: title, authors, abstract, introduction (especially the critiques of previous work in related work / introduction sections).

2. 提取批判链线索

2. Extract critique chain clues

仔细读目标论文的引言和相关工作部分。找出：
  • 它明确说"前人方法 X 有问题 Y"的地方
  • 它声称自己改进了哪篇/哪几篇论文
  • 它对比的 baseline 是谁
从中锁定 被批判/被改进的核心论文（通常 1-3 篇，选最直接的那条线）。
Read the introduction and related work sections of the target paper carefully. Find:
  • Places where it explicitly says "previous method X has problem Y"
  • Which papers it claims to have improved
  • Which baseline it compares against
Lock the core papers that are critiqued/improved from these (usually 1-3 papers, choose the most direct line).

3. 递归溯源（深度研究）

3. Recursive traceability (in-depth research)

对第 2 步找到的核心前序论文，重复同样的过程：它又在批判谁？改进谁？
递归规则：
  • 最多递归 5 层（到第 5 层或到该领域的奠基论文为止）
  • 每层只追 问题最相关的那条线，不发散
  • 如果某层找不到明确的被批判对象，停在那里
使用 Research skill（deep research 模式）获取每层论文的关键信息。每篇论文至少拿到：标题、作者、年份、核心问题、核心解法、对前人的批判点。
Repeat the same process for the core previous papers found in step 2: who are they critiquing? Who are they improving?
Recursion rules:
  • Max 5 recursive layers (stop at layer 5 or the foundational paper of the field)
  • Only follow the most relevant line of the problem for each layer, do not diverge
  • Stop if no clear critiqued target can be found at a layer
Use Research skill (deep research mode) to obtain key information for each layer of papers. For each paper, get at least: title, authors, year, core problem, core solution, critique points of previous work.

4. 前沿延伸

4. Cutting-edge extension

反方向：目标论文之后，有没有新论文在批判/改进它？
同样用 Research skill 搜索：
  • 引用了目标论文的后续工作
  • 同一问题上的最新进展
找到最相关的 1-3 篇后续论文，获取同样的信息。
Reverse direction: after the target paper, are there any new papers critiquing/improving it?
Also use Research skill to search:
  • Follow-up work that cites the target paper
  • Latest progress on the same problem
Find 1-3 most relevant follow-up papers, obtain the same information.

5. 构建演化线

5. Build evolution line

把第 3、4 步的结果整理成时间线：
[最老] Paper_0 → Paper_1 → ... → [目标论文] → [后续论文]
每条箭头标注：后者看到了前者的什么问题。
Organize the results of steps 3 and 4 into a timeline:
[Oldest] Paper_0 → Paper_1 → ... → [Target Paper] → [Follow-up Papers]
Mark each arrow: what problem the latter found in the former.

6. 正向费曼叙事

6. Forward Feynman-style narrative

从最老的论文开始，正向讲述。关键：不是逐篇独立介绍，而是以问题演化为线索串联。
每篇论文讲三件事（以差异为重心）：
  1. 它看到了前人方案的什么具体问题（用例子或场景说明）
  2. 它的解法核心思路（用类比讲清楚）
  3. 这个解法又留下了什么新的问题（自然过渡到下一篇）
Start from the oldest paper, tell the story forward. Key point: do not introduce each paper independently, but connect them with problem evolution as the clue.
Explain three things for each paper (focus on differences):
  1. What specific problem did it find in the previous solution (explain with examples or scenarios)
  2. Core idea of its solution (explain clearly with analogy)
  3. What new problems did this solution leave behind (naturally transition to the next paper)

7. 画图

7. Draw diagrams

两张图：
  • 溯源地图：放在演化叙事之前，展示论文间的引用/批判关系
  • 问题-解法总览：放在叙事之后，把整条线压缩到一屏。让人扫一眼就知道这条线怎么长出来的
Two diagrams:
  • Traceability Map: placed before the evolution narrative, showing the citation/critique relationship between papers
  • Problem-Solution Overview: placed after the narrative, compressing the entire line into one screen. Let people know how the line developed at a glance

8. 提炼洞见

8. Extract insights

读完整条线，回答：
  • 这条演化线背后真正在发生什么变化？（不是表面的技术迭代，是更深层的认知转变）
  • 下一步最可能往哪走？
After reading the entire line, answer:
  • What changes are really happening behind this evolution line? (Not superficial technical iteration, but deeper cognitive shift)
  • What is the most likely next direction?

9. 过红线 + 生成文件

9. Check red lines + generate file

逐条扫红线。额外检查：
  • 因果链条是否连贯——把所有"它看到了什么问题"串起来读，逻辑通不通
  • 差异是否突出——每篇论文的重点是不是在讲"和前面有什么不同"
读 
references/template.org
，按 Denote 规范写入 
~/Documents/notes/
。
Go through the red lines one by one. Additional checks:
  • Is the causal chain coherent — read all "what problem did it find" strings together, is the logic smooth
  • Are differences prominent — is the focus of each paper on "what is different from the previous ones"
Read 
references/template.org
, write to 
~/Documents/notes/
according to Denote specifications.

验收

Acceptance Criteria

  • 问题是主角：读完后记住的是"问题怎么演化的"，不是"有哪些论文"
  • 因果不断：从第一篇到最后一篇，每个转折都有"所以"
  • 差异清晰：每篇论文的独特贡献一句话能说清
  • 外行能跟：不懂这个领域的聪明人读完能复述这条演化线
  • 两张图能独立看：不读正文，只看图也能抓住大意
  • 诚实标注：哪些是确认的引用关系，哪些是推测的，标清楚
  • Problem is the main character: After reading, you remember "how the problem evolved", not "what papers are there"
  • Unbroken causality: From the first paper to the last, every transition has a "so"
  • Clear differences: The unique contribution of each paper can be explained in one sentence
  • Accessible to laymen: Smart people who don't know this field can retell the evolution line after reading
  • Two diagrams are independent: You can grasp the main idea just by looking at the diagrams without reading the text
  • Honest marking: Clearly mark which citation relationships are confirmed and which are speculated