summarise-paper

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

When to use

使用场景

Reading and summarising an academic paper.

读取并总结学术论文。

Workflow

工作流程

Determine whether the input is a PDF file or an arXiv URL.
If the input is a PDF file:
1. Use
```
pdftoppm
```
  to convert the PDF into PNG images. Save them under
```
.cache/{paper name}/
```
  as
```
[1,2,3,...].png
```
  .
2. Read the converted images.
3. You MUST read the paper ONLY via these images (not via text-extraction tools). Papers often contain equations, figures, and charts that must be recognised accurately, and text extraction is unreliable for these.
If the input is an arXiv URL (e.g.
```
https://www.arxiv.org/abs/2601.07372
```
):
1. Normalise the input by replacing the
```
abs
```
  in the URL into
```
src
```
  for TeX source, e.g.:
```
https://www.arxiv.org/src/2601.07372
```
  .
2. Download the source as a local
```
.tar.gz
```
  archive and store it under
```
.cache/downloads/{paper name}/
```
  .
3. Unpack the archive into
```
.cache/{paper name}/
```
  .
4. Locate the entry point (e.g.
```
main.tex
```
  ) and read all relevant contents, including referenced figures.
Summarise the paper as a LaTeX research note and write it under
```
note/{paper name}/
```
, following the Paper Summarisation Instructions below.
The output note must be a standalone LaTeX file that can be compiled directly.

判断输入为PDF文件还是arXiv URL。
若输入为PDF文件：
1. 使用
```
pdftoppm
```
  将PDF转换为PNG图片，保存至
```
.cache/{paper name}/
```
  目录下，命名格式为
```
[1,2,3,...].png
```
  。
2. 读取转换后的图片。
3. 必须仅通过这些图片读取论文内容（禁止使用文本提取工具）。论文中常包含公式、图表等内容，文本提取工具无法准确识别这些元素。
若输入为arXiv URL（示例：
```
https://www.arxiv.org/abs/2601.07372
```
）：
1. 标准化输入URL，将其中的
```
abs
```
  替换为
```
src
```
  以获取TeX源文件，示例：
```
https://www.arxiv.org/src/2601.07372
```
  。
2. 将源文件下载为本地
```
.tar.gz
```
  压缩包，存储至
```
.cache/downloads/{paper name}/
```
  目录下。
3. 将压缩包解压至
```
.cache/{paper name}/
```
  目录。
4. 找到入口文件（如
```
main.tex
```
  ）并读取所有相关内容，包括引用的图表。
按照下方的《论文总结规范》，将论文总结为LaTeX格式的研究笔记，保存至
```
note/{paper name}/
```
目录下。
输出的笔记必须为可直接编译的独立LaTeX文件。

Paper Summarisation Instructions

论文总结规范

You are a senior researcher searching for new ideas for your next top-tier conference/journal paper. You read papers and summarise them into notes. Your goal is to produce a “weeks-later readable” research note: after reading many papers, I should be able to reconstruct the paper’s core ideas, methods, and evidence, and discuss it intelligently even if I have not read it recently.

你是一名资深研究员，正在为下一篇顶级会议/期刊论文寻找新思路。你需要阅读论文并将其整理为研究笔记。目标是撰写一份「数周后仍能快速理解」的研究笔记：在阅读大量论文后，即使我近期未重读该论文，也能通过这份笔记重构论文的核心思想、方法和论据，并进行有条理的讨论。

Core requirements

核心要求

Faithfulness: Use ONLY information supported by the paper. If a detail is missing or unclear, write “Not stated in the paper” (or “Not shown in the provided excerpt”) rather than guessing.
Grounding: Where possible, cite the source of each key claim (section name, figure/table number, equation number, or page number).
Clarity: Prefer intuition-first explanations, then formalisation/maths, then implications.
Completeness: The Methodology section must be self-contained and read as a coherent story from inputs → computations → outputs, including training and inference pipelines if applicable.
Language: You must use British English.

忠实性： 仅使用论文中明确提及的信息。若某一细节缺失或表述模糊，需注明「论文中未提及」（或「提供的摘录中未展示」），禁止主观猜测。
溯源性： 尽可能为每个关键论点标注来源（如章节名称、图表/表格编号、公式编号或页码）。
清晰性： 优先采用直觉化解释，再进行形式化表述/数学推导，最后说明研究意义。
完整性： 方法论部分需自成体系，连贯讲述从输入→计算→输出的完整流程，若涉及训练和推理 pipeline 也需包含在内。
语言： 必须使用英式英语。

Notation rules

符号规则

$\mathcal{C}$ denotes a set.
Bold lowercase $\mathbf{x}$ denotes a vector; bold uppercase $\mathbf{X}$ denotes a matrix.
Uppercase $X$ denotes a random variable; lowercase $x$ denotes a deterministic value.
Use correct LaTeX ($...$ for inline maths, $$...$$ for display maths). Define symbols before using them.

Write the note using the EXACT structure and headings below:

$\mathcal{C}$ 表示集合。
粗体小写 $\mathbf{x}$ 表示向量；粗体大写 $\mathbf{X}$ 表示矩阵。
大写 $X$ 表示随机变量；小写 $x$ 表示确定值。
使用标准LaTeX格式（行内公式用$...$，块级公式用$$...$$）。符号需先定义再使用。

请严格按照以下结构和标题撰写笔记：

1. Motivation

1. 研究动机

What problem is being addressed?
What failure mode or limitation of prior work is targeted?
Why it matters.

论文旨在解决什么问题？
针对现有研究的哪些缺陷或局限性？
该研究的重要性体现在何处？

2. Contributions

2. 研究贡献

Bullet list of the paper’s concrete contributions (methods, theory, benchmarks, analyses).
Where possible, separate contributions into “new idea” vs “engineering/implementation” vs “evaluation/protocol”.

以项目符号列表形式列出论文的具体贡献（包括方法、理论、基准测试、分析结果等）。
尽可能将贡献分为「创新思想」「工程实现」「评估方案」三类。

3. Methodology

3. 研究方法

MINI-PAPER STYLE, single coherent story.
Write this section as a flowing narrative (like the Methods section of a well-written paper), not as a report or checklist.

采用「迷你论文」风格，讲述连贯完整的故事。
本部分需以流畅的叙事方式撰写（参考优质论文的方法章节），而非报告或清单形式。

Hard constraints

硬性约束

No sub-bullets or lettered substeps. Minimal headings are allowed, but prefer continuous prose.
The story must flow in one direction: introduce concepts only when they become necessary.
Each paragraph should lead naturally into the next (use transitions such as “To address this…”, “Concretely…”, “This enables…”, “At inference time…”).
If a critical detail is missing, explicitly write “Not stated in the paper” rather than guessing.

禁止使用子项目符号或字母编号的子步骤。可使用少量标题，但优先采用连续段落。
叙事需遵循逻辑顺序：仅当概念必要时再引入。
段落之间需自然过渡（可使用「为解决这一问题…」「具体而言…」「这一设计使得…」「推理阶段…」等衔接语）。
若关键细节缺失，需明确标注「论文中未提及」，禁止主观猜测。

Narrative guidance

叙事指南

Present the method in the order the paper itself develops it (often: problem → idea → formulation → algorithm → training → inference → complexity).
Include whichever of the following are relevant, but do not force all items or a fixed order:
- Problem setting and assumptions (inputs, outputs, constraints)
- Core insight and how it differs from prior work (if discussed)
- The main objects/components (model modules, memory, prompts, optimiser, data stream, etc.)
- Key equations/objectives (only if present; define symbols before use)
- The end-to-end procedure (including the training loop if applicable)
- Inference-time behaviour (if different from training or otherwise non-trivial)
- Notable implementation details or computational overheads (only if stated)

按照论文自身的逻辑顺序呈现方法（通常为：问题→思路→形式化→算法→训练→推理→复杂度分析）。
包含以下相关内容，但无需强制覆盖所有项或固定顺序：
- 问题设定与假设（输入、输出、约束条件）
- 核心创新点及其与现有研究的差异（若论文有所讨论）
- 核心组件/对象（如模型模块、内存、提示词、优化器、数据流等）
- 关键公式/目标函数（仅当论文中存在时需包含，符号需先定义再使用）
- 端到端流程（若涉及训练循环也需包含）
- 推理阶段的行为（若与训练阶段不同或具有特殊性）
- 值得注意的实现细节或计算开销（仅当论文中提及）

4. Experimental Setup

4. 实验设置

A short “Implementation checklist” with 4–8 items ONLY if the paper provides those details (e.g., buffer size, optimiser, key hyperparameters, compute, architectural choices). If not provided, write “Not stated in the paper.”
Datasets/tasks, baselines, evaluation protocol, and metrics.
What ablations or sensitivity analyses were run.

若论文提供相关细节，需列出4-8项的「实现检查清单」（如缓冲区大小、优化器、关键超参数、计算资源、架构选择等）。若未提供，需注明「论文中未提及」。
数据集/任务、基线模型、评估方案及指标。
开展了哪些消融实验或敏感性分析。

5. Strengths & Weaknesses

5. 优势与不足

Strengths: 3–6 bullets with reasons tied to evidence in the paper.
Weaknesses/risks: 3–6 bullets (e.g., missing baselines, unclear protocol, confounders, scaling limits, assumptions, failure cases).
Include “What I would ask the authors as a reviewer” (2–3 questions).

优势： 列出3-6个项目符号，每个优势需结合论文中的证据说明原因。
不足/风险： 列出3-6个项目符号（如缺失基线模型、方案表述模糊、混杂变量、缩放限制、假设条件、失效场景等）。
需包含「作为审稿人，我会向作者提出的问题」（2-3个问题）。

6. Final short note

6. 最终简短总结

1–3 sentences giving a crisp description of what the paper does and the key result/claim (avoid numbers unless explicitly stated in the paper).
Key takeaways:
- 3–5 bullets covering what I should steal/adapt, what to be cautious about, and one concrete follow-up experiment idea.
- This section may be opinionated, but it must not invent facts about the paper.

用1-3句话简洁描述论文的核心工作及关键结论/主张（除非论文明确提及，否则避免使用具体数字）。
关键要点：
- 列出3-5个项目符号，包括可借鉴/改编的内容、需谨慎对待的部分，以及一个具体的后续实验想法。
- 本部分可包含主观观点，但不得编造论文中未提及的事实。

Dependencies (install if missing)

依赖项（缺失时请安装）

bash

undefined

bash

undefined

macOS (Homebrew)

brew install poppler

Ubuntu/Debian

sudo apt-get install -y poppler-utils


If installation isn't possible in this environment, tell the user which dependency is missing and how to install it locally.

sudo apt-get install -y poppler-utils


若当前环境无法安装，请告知用户缺失的依赖项及其本地安装方法。