Arxiv Paper Reading

Always convert the paper to Markdown before analysis.

1. Normalize input to an arXiv identifier.
   • Example: 2509.16117 is the arXiv ID for https://arxiv.org/abs/2509.16117.
   • abs, pdf, and html URLs map to the same ID.
2. Use local conversion with uvx first:

• uvx --from arxiv2markdown arxiv2md <ARXIV_ID> --remove-refs -o {paper-title}-{arxiv-id}.md

Preferred example:

uvx --from arxiv2markdown arxiv2md 2509.16117 --remove-refs -o title_2509.16117.md

Do not use python libraries or other tools for fetching/conversion to avoid inconsistencies. Prefer uvx --from arxiv2markdown arxiv2md for this step.

Fallback Rules

• If uvx is unavailable or conversion fails, call arxiv2md API with curl:
  • curl -s "https://arxiv2md.org/api/markdown?url=<ARXIV_ID>" > paper.md
• Then read and analyze from paper.md directly.
• Example fallback command:
  • curl -s "https://arxiv2md.org/api/markdown?url=2509.16117" > paper.md

Reading Workflow

1. Locate key sections in order: Abstract, Introduction, Method, Experiments, Limitations/Conclusion.
2. Extract explicit evidence lines for each claim before writing conclusions.
3. Distinguish paper claims from your own inference.
4. If a requested detail is missing, say it is not specified instead of guessing.

Output Template

Use this exact section order unless the user asks otherwise.

1) 一句话核心贡献

• One sentence only.

2) 问题背景

• 要解决什么问题
• 之前方法的主要不足
• 作者为什么认为他们的方法可行

Cover each item with evidence-oriented explanation. Prefer short paragraphs; use bullets only if comparison becomes clearer.

3) 方法详解

• 模型架构
  • Backbone 是什么
  • 基于什么已有模型/范式改进
• 新增模块逐项拆解
  • 动机
  • 具体实现
  • 与原模型交互方式

For each module, prefer this format:

• 模块名
• 动机
• 实现
• 代价/复杂度影响

When writing, avoid turning every module into a flat checklist; start with a concise intuition paragraph, then provide the structured fields.

4) 批判性思考

• 最大亮点
• 局限性
• 可行改进方向
• 对 community 的价值与 high-level insight

Separate "paper-supported" and "your inference" when needed.

Quality Bar

• Do not fabricate numbers, datasets, or ablation results.
• When citing performance, include metric + benchmark/task context.
• Mark uncertainty explicitly if table/figure parsing is ambiguous.
• Be detailed by default, with enough context that a graduate-level reader can follow the argument without reopening the paper.