paper-repro-python（paper repro Python 搜索）

Follow this workflow end-to-end unless the user explicitly asks to skip steps

1) Intake and scope

• Confirm input artifacts: TeX source path(s), PDF path, supplementary files, target repository, and expected outputs.
• State assumptions explicitly when information is missing.
• Keep approach adaptable to the specific paper; do not force a fixed dependency stack or rigid project template.
• Check whether the working folder already contains paper source files (.tex, .bib, style files, figures).
• Check whether the working folder contains user-preprocessed documents (.md, .json, images such as .png, .jpg, .svg).
• Source priority rule (read in order, stop when sufficient):

1. TeX sources (preferred): If usable TeX source files (.tex, .bib, style files) are present, use them as the primary source. 2. User-preprocessed documents (secondary): If TeX is absent or incomplete, read user-provided documents (.md, .json) and images (.png, .jpg, .svg) that may contain pre-extracted paper content. 3. PDF fallback (last resort): Only when both TeX and user-preprocessed documents are unavailable or insufficient, fall back to PDF extraction.

2) Source extraction (TeX → preprocessed docs → PDF)

• TeX path (highest priority):

- Parse and read the main TeX project structure first (main.tex or equivalent entry file and includes). - Preserve original scientific wording when converting relevant content to Markdown notes. - Resolve equations, theorem blocks, citations, and appendices from source files whenever possible. - Record unresolved include/bibliography issues explicitly; do not invent missing content.

• User-preprocessed documents path (secondary):

- Read Markdown files (.md) that may contain paper content extracted by the user. - Read JSON files (.json) that may contain structured paper data (metadata, sections, references). - View images (.png, .jpg, .svg) that may contain paper figures, tables, or scanned pages. - Preserve original content; do not summarize or paraphrase. - Note the source of each piece of information (which file, which section).

• PDF fallback path (lowest priority, when all else fails):

- Extract paper content page by page into Markdown, preserving the original wording. - Do not summarize, paraphrase, or rewrite scientific statements. - Preserve structure faithfully: - Title, authors, affiliations, abstract, sections, subsections. - Equations (LaTeX-friendly when possible), theorem/lemma/proposition blocks. - Tables, figure captions, references, appendices, footnotes. - If a PDF is scanned or partially unreadable: - Run OCR and mark uncertain spans clearly. - Never silently invent missing text. - Include image references/placeholders when figures cannot be represented as plain text. - Produce one primary output file such as paper_fulltext.md.

3) Extraction quality checks

• Validate completeness before moving to reproduction:

- Section/headings coverage matches the TeX project or PDF source used. - Key equations and algorithm blocks are present. - References and appendices are included if present in the source.

• Report known extraction limitations and exact affected files/pages/segments.

4) Reproduction planning (paper-specific)

• Build a reproduction plan from the extracted source materials (TeX-derived notes and/or Markdown), not from memory.
• Identify:

- Problem definition, notation, assumptions, and objective functions. - Algorithm steps and required components. - Dataset generation/loading, training/optimization, and evaluation protocol. - Baselines and ablations required for faithful reproduction.

• If details are missing or ambiguous, call out the gap and provide a conservative implementation choice with rationale.

5) Python implementation principles

• Implement with modular design and clear boundaries:

- Separate concerns (data, models/algorithms, training/solver loop, evaluation, utils, config). - Prefer low coupling and high cohesion.

• Avoid monolithic scripts:

- Split code into modules whenever responsibilities can be separated. - Prefer one clear responsibility per file.

• File size guideline:

- Keep a single source file under ~200 lines whenever practical. - If a file grows beyond ~200 lines, refactor into submodules unless there is a clear reason not to.

• Keep dependencies minimal and paper-driven; choose tools based on the paper's actual needs.
• Avoid over-engineering early; start from the minimum reproducible core, then extend.
• Add tests/checks for critical math or pipeline steps where feasible.
• Preserve reproducibility:

- deterministic seeds when applicable, - explicit config for key hyperparameters, - clear experiment entry points.

6) Logging and data persistence

• All reproduction experiments must save execution logs and output data:

- Logs: Save console output, training progress, and error messages to timestamped log files (e.g., logs/experiment_YYYYMMDD_HHMMSS.log). - Data: Save all generated data (metrics, model checkpoints, intermediate results) to structured files (e.g., JSON, CSV, or pickle) in a dedicated outputs/ or results/ directory. - Configuration snapshot: Save the exact configuration/hyperparameters used for each run alongside the outputs.

• Logging should capture enough detail to:

- Trace the execution flow for debugging. - Compare results across multiple runs. - Identify the exact conditions under which results were produced.

7) Result verification and comparison

• After running reproduction experiments, compare results against the paper's reported data:

- Extract quantitative metrics (tables, figures) from the paper source. - Compute the same metrics from reproduction outputs. - Document both paper-reported values and reproduction values side by side.

• Identify discrepancies:

- If reproduction results deviate significantly from paper results, investigate potential causes: - Implementation errors (algorithm bugs, missing steps). - Hyperparameter differences. - Random seed or initialization differences. - Dataset or preprocessing differences. - Document all findings and fixes applied.

• Validation criteria:

- Define acceptable tolerance for numerical differences based on the paper's domain. - Flag results outside tolerance for further review.

• The goal of reproduction is accurate reconstruction of paper results; verification is mandatory, not optional.

8) README header requirements (paper metadata)

• Every reproduction project README must start with paper metadata before any other content:

- Paper title (original title as published) - Authors (full names, affiliations, and email addresses if available) - Abstract (verbatim copy of the original abstract)

• For README_zh-CN.md:

- Paper title: provide Chinese translation if original is in English; keep original if paper is in Chinese. - Authors: keep original names and affiliations; translate country/region names if needed. - Abstract: provide faithful Chinese translation of the abstract.

• Format example (English README):

  # [Paper Title]

  **Authors:** Author Name¹, Co-Author Name²
  **Affiliations:**
  ¹ Department, University, Country (email@university.edu)
  ² Lab, Institution, Country (email@institution.edu)

  ## Abstract

  [Verbatim abstract text from the paper]

  ---

  [Then reproduction project content begins...]

9) README update requirements (bilingual + images)

• Generate and maintain two README files after code changes:

- README.md (English original) - README_zh-CN.md (Chinese translation aligned with the English version)

• After the paper metadata header, ensure both files include:

- paper citation and target claims to reproduce, - environment/setup commands, - project structure overview and module responsibilities, - how to run main experiments, - expected outputs/metrics and where artifacts are saved, - known deviations from the paper and why.

• Insert generated figures/images into both README files using valid relative Markdown image paths.
• Image output granularity rule: unless multi-panel comparison is explicitly needed, save one chart per image file (one figure per file).
• Keep both README files aligned with actual code paths and commands.
• Keep Chinese content as faithful translation of English technical content (no missing key steps).

10) Output contract

• Deliver:

- source-derived extraction notes/file(s) (TeX-first, PDF fallback when needed), - implemented/updated Python code, - README.md and README_zh-CN.md with embedded generated images.

• Clearly separate:

- exact extracted content (verbatim from source), - your implementation notes and engineering decisions.

• Report reproduction status:

- which claims/experiments were successfully reproduced, - known gaps or deviations from paper results, with reasons.