Li Mu Paper Reading Notes

Problem

Paper reading often feels overloaded: it is hard to tell whether a paper deserves close reading, or to separate claims, evidence, assumptions, and reproduction gaps quickly. This skill applies Li Mu's paper-reading method to turn a paper into a structured note that supports judgment, review, comparison, and implementation planning, rather than a generic summary.

Core workflow

• Input handling: accepts arXiv links, PDF links, local PDFs, DOIs, and titles; when no PDF is available, it can use Mermaid architecture diagrams instead of extracted figures.
• Three-pass reading: creates a Paper Card to decide whether to continue, then maps the problem, method, experiments, and evidence, and finally performs a virtual reproduction pass to surface assumptions, steps, missing details, and risks.
• Argument audit: tracks the chain from problem to claim to reason to evidence to limitation, labels evidence as strong, weak, or unsupported, and checks for data leakage, unfair baselines, metric mismatch, and omitted negative results.
• HTML output: generates a self-contained HTML note with embedded base64 images, MathJax-rendered formulas, and a local browser view; useful for close reading, review, comparison, reproduction planning, and literature mapping.

Boundaries

Close reading is strongest when the full paper is accessible. Title-only runs rely on retrieved abstracts or text and may be less stable than PDF-based analysis. Figure extraction depends on the PDF and image parsing; without figures, schematic diagrams may replace them. HTML formula and diagram rendering rely on CDN assets, so offline use may show raw LaTeX or fail to render diagrams.