deep-paper-reading

name: deep-paper-reading description: Deeply read, investigate, critique, and write a detailed Chinese Markdown report for academic papers. Use when the user asks to analyze a paper, PDF, arXiv article, conference paper, journal article, technical report, or manuscript in depth, especially when they request a saved md report, Chinese literature overview, author background, Introduction and Related Work investigation, detailed method breakdown, framework analysis, detailed experiment analysis, code or pretrained weight availability, limitations, insights, or future research directions.

Deep Paper Reading

Overview

Use this skill to produce a deep paper-reading report rather than a shallow abstract summary. The report must combine paper-internal evidence with external investigation for citation metadata, author background, related methods, code availability, pretrained weights, and broader research context. The final deliverable should be a saved Markdown file unless the user explicitly asks for chat-only output.

Hard language requirement: the saved Markdown report MUST be written in Chinese by default. Do not write the main prose, section explanations, critique, experiment analysis, limitations, or conclusion in English unless the user explicitly asks for an English report. English is allowed only for exact paper titles, method names, module names, dataset names, metric names, code identifiers, URLs, citations, and short technical terms whose Chinese translation would reduce precision.

Output Artifact

1. Save the complete report as a .md file by default.
2. If the user gives an output path, write the report there.
3. If no output path is given, create a descriptive filename in the current workspace using the paper title, for example deep_reading_<short-paper-title>.md.
4. In the chat response, provide only a concise completion summary and a link to the saved Markdown file.
5. Do not stop at a brief inline summary when the user asks for deep reading. The Markdown report is the primary deliverable.
6. The saved Markdown report must be Chinese-first: Chinese title, Chinese section headings, Chinese table captions, Chinese figure captions, Chinese analysis paragraphs, and Chinese final summary.
7. Include a Chinese source list or references section in the Markdown file with links for web-verified information.
8. If some required information cannot be found, write 未找到可靠公开信息 or 论文未报告 in the report rather than omitting the item.
9. Create an asset directory next to the report when inserting paper figures, for example deep_reading_<short-paper-title>_assets/.
10. Save extracted or rendered paper figures into that asset directory and insert them into the Markdown file with image syntax.

Source Handling

1. Prefer the official paper PDF, arXiv PDF, publisher page, project page, and official code repository.
2. When the user provides only a title, DOI, arXiv ID, URL, screenshot, or local PDF, identify the canonical paper first.
3. Browse the web for information that may change or requires public verification, including Google Scholar-style citation, author affiliations, author research background, code availability, pretrained weights, project pages, and repository status.
4. Cite sources with links when using external information.
5. Separate paper facts from interpretation. Mark uncertain or inferred statements explicitly.

Language Style

1. Write the Markdown report in Chinese by default. Treat Chinese output as mandatory unless the user explicitly requests another language.
2. Keep established English technical terms, component names, method names, dataset names, metric names, model names, and paper titles in English when translating them would reduce precision.
3. Avoid unnecessary Chinese-English mixing in ordinary prose. Prefer Chinese verbs, connectors, and explanations around English terms.
4. Translate generic labels into Chinese when possible, for example use 问题, 动机, 局限性, 实验设置, 评价指标, 消融实验, 可视化验证.
5. Preserve exact names such as Segment Anything Model, SAM, Dual Stream Adapter, Bidirectional Knowledge Distillation, Mixed Prompt Embedding, COD10K, F-measure, and S-measure.
6. When introducing an English term for the first time, give a short Chinese explanation if useful. After that, use the English term consistently.
7. Do not leave template labels such as Problem:, Motivation:, Limitations:, Method overview:, Experiment setup:, or Conclusion: in English. Translate them to Chinese labels such as 问题：, 动机：, 局限性：, 方法总览：, 实验设置：, and 总结：.
8. Before finishing, scan the saved Markdown for long English prose paragraphs. If any are found, rewrite them in Chinese while preserving exact technical names.

Figure Handling

1. Insert the paper's original figures in the Markdown report when explaining method and experiment results.
2. For Step 3, include the original architecture, pipeline, module, algorithm, or framework figures that are needed to understand the method.
3. For Step 4, include the original qualitative comparison, visualization, ablation visualization, or result figures that are needed to understand experimental evidence.
4. Prefer extracting the original figure image from the PDF. If exact figure extraction is impractical, render or screenshot the relevant PDF page region and crop around the figure.
5. Save figures with stable names, for example fig2_pipeline.png, fig3_adapter.png, fig5_qualitative.png.
6. Insert each figure near the paragraph that explains it, not all figures at the end.
7. Add a short caption below the image stating the original figure number and what the reader should notice.
8. Do not redraw the paper's original figure as Mermaid when the original figure is available. Use Mermaid only as an additional reconstruction if the original figure is missing, unreadable, or useful for simplifying a complex pipeline.
9. If a required figure cannot be extracted, write 原图未能从 PDF 中可靠提取 and still explain the figure from the paper text.

Main Workflow

Follow these five steps unless the user asks for a shorter or different format. Use Chinese section titles by default, for example 第一部分：文献总览, 第二部分：背景信息调研, 第三部分：核心方法详解, 第四部分：实验与结果, and 第五部分：总结.

Step 1: 文献总述

Produce a complete overview:

1. Provide the full citation in a style consistent with Google Scholar when available. If Google Scholar is inaccessible, use the paper's official citation, publisher metadata, DBLP, Semantic Scholar, OpenAlex, or arXiv metadata, and state the source.
2. State the paper title and the research field or subfield.
3. Identify the first author and corresponding author. Investigate their affiliations, research directions, representative background, lab or homepage, and other public information relevant to understanding the paper.
4. Answer these questions:
   • What problem does the paper solve?
   • What is the research motivation?
   • What limitations exist in current models or methods for this task?
   • What are the main innovations?
   • What performance or effects does the paper achieve?
   • Is the code open source?
   • Are pretrained weights, checkpoints, models, datasets, or demos provided?

Step 2: 背景信息调研

Investigate the background around Introduction and Related Work in detail. This section should read like a compact literature review, not a short list:

1. Extract the key methods, model families, datasets, tasks, and technical lines mentioned in Introduction and Related Work.
2. Download or open the most relevant cited papers when needed, prioritizing papers that define the task, introduce major baselines, or represent the closest prior work.
3. Classify related work by the themes used in the paper when possible. If the paper's categories are weak, create a clearer taxonomy and explain the mapping.
4. For each theme, summarize the development trajectory:
   • Early or foundational methods.
   • Representative model families.
   • Recent progress.
   • Persistent bottlenecks.
   • How the target paper positions itself in this theme.
5. For each important cited paper, explain why it matters to the target paper. Do not merely list paper names.
6. At the end of Step 2, add a subsection titled 本节提到的文献 and list the cited papers mentioned in the background investigation.
7. Each reference entry should include available bibliographic information such as title, authors, venue/year, and a short note about its role in the discussion.
8. Read references/related-work-survey.md when producing a detailed related-work investigation.

Step 3: 核心方法详解

用中文从系统层面到组件层面解释方法。本节应是报告中最长、最详细的部分之一：

1. 先写方法总览：
   • 总体网络架构。
   • 主要模块或组件。
   • 训练阶段数据流。
   • 推理阶段数据流。
   • 输入、输出、损失、目标和监督信号。
2. 详细介绍框架图：
   • 定位论文中的主架构图。
   • 将原始框架图插入 Markdown 报告。
   • 解释每个 block、arrow、branch、tensor 或 stage 的含义。
   • 描述数据如何从输入流向输出。
   • 将图中标签对应到方法文字和公式。
   • 如果图缺失或不清楚，用中文文字或 Mermaid 重建流程。
3. 按顺序解释每个主要组件：
   • 设计目的。
   • 输入和输出。
   • 内部操作。
   • 相关公式或算法。
   • 论文报告或可推断的 tensor shape、feature dimension、resolution change。
   • 可训练参数与冻结参数。
   • 与原始 baseline 方法的关系。
   • 为什么需要该组件。
   • 如果移除或替换该组件会发生什么，并结合消融证据说明。
4. 对每个公式说明：
   • 每个变量的含义。
   • 正在执行的操作。
   • 为什么包含该公式。
   • 它如何影响训练、推理或表征学习。
5. 对每个算法设计选择，说明作者动机、预期收益和可能弱点。

Step 4: 实验与结果

用中文把实验当作证据来分析，而不是装饰性复述。本节应足够详细，使读者理解如何复现和评价该工作：

1. 解释实验设置：
   • 训练配置。
   • 推理配置。
   • 论文报告的硬件或计算资源。
   • 超参数、预处理、数据增强、优化器和停止准则。
2. 列出每个数据集。对每个数据集介绍：
   • 任务与领域。
   • 数据规模。
   • 训练、验证和测试划分。
   • 标注类型。
   • 在该领域中的常见用途。
3. 列出对比方法。介绍每个 baseline 或 baseline family，解释为什么相关，并说明它是经典方法、最新 SOTA、foundation-model 方法、同模态方法、跨模态方法还是直接前作。
4. 列出评价指标。解释每个指标衡量什么、越高/越低是否更好、为什么适合该任务、有什么局限。
5. 分析消融实验：
   • 消融了哪些组件。
   • 每个消融回答什么问题。
   • 消融是否真正支持论文声称的贡献。
   • 重要行的精确指标变化。
   • 消融是否隔离了单一因素，还是混合了多个因素。
   • 为了更强证据还缺少哪些消融。
6. 分析可视化或定性验证：
   • 可视化了什么。
   • 在可用时插入原始可视化或定性对比图。
   • 支持哪条主张。
   • 样例是否具有代表性，是否可能 cherry-picked。
   • 展示或没有展示哪些失败模式。
7. 只重构紧凑核心结果表。保留关键结果、关键对比和关键消融的精确数字，但不要在报告中复制过宽或过长的完整表格。
8. 按数据集和指标，将本文方法与最强 baseline 比较，不要只与弱 baseline 比较。
9. 讨论统计可靠性：多随机种子、方差、置信区间、测试协议、数据泄漏风险，以及论文是否报告这些信息。
10. 需要详细实验审查时，读取 references/experiment-analysis-checklist.md。

Step 5: 总结

用中文写出有解释力、且基于证据的总结：

1. 说明模型局限，包括作者承认的局限，以及从方法或实验中推断出的局限。
2. 解释该工作的启发：
   • 概念启发。
   • 技术设计启发。
   • 实验设计启发。
   • 应用启发。
3. 提出受该工作启发的潜在研究方向：
   • 直接扩展。
   • 更强 baseline 或评价。
   • 新数据集或新任务。
   • 模型压缩、鲁棒性、泛化性、可解释性、部署或跨域迁移方向。

Output Standards

1. Use the report structure in references/report-template.md for the saved Markdown report.
2. Use clear headings and preserve the five-step order.
3. Write a complete Chinese report, not an outline. Avoid overly terse bullets in Step 3 and Step 4; use Chinese paragraphs, tables, and itemized explanations.
4. Make Step 2 detailed enough to explain the field's development. End Step 2 with a reference list of the papers discussed in that section.
5. Insert original paper figures in Step 3 and Step 4 whenever they are central to the explanation.
6. Keep reconstructed experiment tables compact: include only the proposed method, the strongest direct baselines, important foundation-model baselines, and the most important metrics.
7. Use Chinese as the required report language while preserving precise English names for methods, modules, datasets, metrics, and paper titles.
8. Do not rely only on the abstract. Verify claims against method, experiment tables, ablations, and appendices.
9. Do not copy the authors' framing uncritically. For each major claim, ask what evidence supports it.
10. Use external sources for author background, code status, pretrained weights, and related-paper context.
11. Include source links for external findings.
12. Clearly state when a detail is not found in the paper or public sources.
13. End the Markdown report with an evidence table mapping claims to supporting paper sections, tables, figures, or external sources.

Resource Guide

• references/report-template.md: Use for the final full report structure.
• references/related-work-survey.md: Use when investigating Introduction and Related Work in detail.
• references/experiment-analysis-checklist.md: Use when auditing experiments, ablations, metrics, and visualization evidence.
• scripts/extract_pdf_text.py: Use for local PDF text extraction when direct PDF reading is unreliable and PyMuPDF is available.