---

name: paper-reviewer description: Review research papers (especially PDFs). Use when the user asks to read/通读/讲解/总结/审稿 a paper and wants a Chinese-first explanation of what it does, what is novel (创新点), plus reviewer-style strengths/weaknesses, major/minor concerns, and questions to authors.

Paper Reviewer

Overview

Read a paper end-to-end (prefer PDF), then produce a teachable explanation and a reviewer-style critique: content summary, innovation points, evidence quality, and actionable concerns.

Quick Start (Inputs)

• Paper: local PDF path (preferred), or arXiv/DOI/citation.
• Audience: beginner / familiar-with-field / expert.
• Focus: method / experiments / critique / implementation.
• Depth: 10-min / 30-min / 90-min talk notes (default: 30-min).
• Target venue (optional): e.g., NeurIPS/ICLR/ACL, or "internal reading group".

If the user does not specify, assume: audience="熟悉基础 ML", focus="method + experiments + critique", depth="30-min", language="Chinese".

Workflow

1) Identify the paper

• If multiple PDFs exist, ask which one to review.
• Record title/authors/venue/year (as shown), and page count.

2) Extract text and render pages (prefer visual skim)

• Use the helper script to extract per-page text and (optionally) render pages to PNG for figure/table inspection:
  • python3 skills/paper-reviewer/scripts/dump_paper_pdf.py --pdf "<PATH>" --out-dir "tmp/paper-review/<slug>" --render
• If rendering fails (missing fitz/PyMuPDF), rerun without --render and continue.

3) First pass: map the paper (10-20 min)

• Identify:
  • Problem setting, inputs/outputs, assumptions.
  • 3-5 core contributions (claimed novelty).
  • The "main loop" of the method in one paragraph.
  • Which experiments are intended to support which claims.

4) Second pass: teach the method

• Explain in this order (even if the paper orders differently):
  1. Problem + why it matters.
  2. Baseline mental model (what a reasonable approach would do).
  3. What is new (the delta vs baselines/prior work).
  4. Method (step-by-step; pseudocode-level).
  5. Complexity and failure modes.
• For equations: explain what each term does, not just restate symbols.
• When referencing results, cite section/figure/table numbers (and page numbers if helpful).

5) Third pass: experiments and evidence

• For each experiment:
  • State the claim being tested.
  • Describe the setup (data, metrics, protocol, baselines).
  • Interpret the result: what it supports; what it does not.
  • Call out confounds: data leakage, unfair tuning, missing ablations, weak baselines, small sample, cherry-picking.

6) Innovation analysis (创新点核验)

• For each claimed innovation, answer:
  • What is new?
  • Why does it matter (what capability improves)?
  • What prior work is it closest to (most plausible "already known" baseline)?
  • What evidence supports the claim?
  • What experiment/ablation would falsify it?

7) Reviewer-style critique

• Use references/review_rubric.md as a checklist.
• Avoid long verbatim quotes; paraphrase.

Output Format (Recommended)

• 一句话结论 (TL;DR)
• 这篇论文在做什么 (problem + setting)
• 方法概览 (core idea + main loop)
• 贡献/创新点 (3-6 bullets; claimed vs assessed)
• 方法细讲 (module-by-module; pseudocode-level)
• 实验解读 (what each supports)
• 优点 (Strengths)
• 主要问题 (Major concerns) (actionable: why it matters + what to add/change)
• 次要问题 (Minor concerns)
• 给作者的问题 (questions to clarify)
• 可复现性清单 (data/code/hparams/eval)
• If asked: 评分 + 信心

Script: scripts/dump_paper_pdf.py

• Purpose: extract per-page text and (optionally) render pages to PNGs.
• Outputs (under --out-dir):
  • metadata.json
  • text_by_page.txt
  • headings_guess.txt
  • captions_guess.txt
  • render/ (PNG pages if --render)