Translate Paper Zh

Required Output Layout

Keep one paper per subfolder under the user's paper/ directory.

Input staging is allowed:

• the user may initially download a raw paper PDF and/or raw source folder directly into paper/
• treat those root-level items as temporary intake material, not as the final organized layout
• before finishing, normalize everything into one dedicated paper subfolder and remove duplicate root-level intake copies when it is safe to do so

Preferred layout:

paper/
  <paper-slug>/
    README.md
    <paper-slug>_en.pdf
    <paper-slug>_cn.pdf
    <source-dir>/
    <source-dir>_cn/

Rules:

• Choose a concise folder slug from the paper title, topic, or arXiv id.
• Default to lowercase hyphen-case slugs such as llm-agent-survey-2503-21460v1.
• Name the root-level PDFs exactly <paper-slug>_en.pdf and <paper-slug>_cn.pdf.
• Preserve the original English source tree unchanged inside <source-dir>/.
• Place the translated LaTeX tree in a sibling directory named <source-dir>_cn/.
• If the user only provides loose files, first create the paper subfolder and normalize the files into this layout.
• After normalization, the paper/ root should mainly contain shared folders plus one subfolder per paper, not duplicate loose PDFs or loose source trees for papers that have already been organized.
• Write README.md in Chinese with: paper title, one-paragraph summary, main contributions, method sketch, tags, and a short file map.

Workflow

1. Inventory and normalize the intake

First inspect the user's paper/ directory and identify which items belong to the target paper:

• a loose PDF such as 2503.21460v1.pdf
• a loose source tree such as arXiv-2503.21460v1/
• an already normalized paper folder

If the paper is not yet normalized:

• choose the target <paper-slug>/
• move or copy the loose English PDF into <paper-slug>/<paper-slug>_en.pdf
• move or copy the loose English source tree into <paper-slug>/<source-dir>/
• do all later work inside that paper subfolder

Prefer moving over copying when the loose root-level items are clearly just staging inputs for this paper and are not referenced elsewhere by the user. If there is any doubt, keep the source, finish the translation, then ask before deleting ambiguous leftovers.

2. Inventory the source tree

Run:

python3 scripts/collect_tex_tree.py /path/to/paper/<source-dir>

Use the output to identify:

• the main .tex entrypoint
• included chapter or section files
• bibliography files
• major figure assets

If there are multiple main-file candidates, prefer the one that contains the complete paper body and bibliography.

3. Create the translated source tree

Copy the English source directory to a sibling _cn directory before translating. Make all translation edits only inside the _cn tree.

Preserve these items unless compilation forces a minimal localized change:

• macro names
• labels and reference keys
• citation keys
• filenames and relative paths
• package names and class names
• equations and math delimiters

If Chinese support is missing, patch only the translated copy's preamble to keep XeLaTeX compilation working. Do not rewrite unrelated formatting.

Preferred Chinese compilation strategy:

• prefer ctex with fontset=none
• prefer installed font family names first; switch to explicit font file paths only when family resolution is unreliable
• use a serif body font for papers when available, for example Source Han Serif SC, Noto Serif CJK SC, or Songti SC
• use a clean sans fallback for headings or UI-like text, for example Hiragino Sans GB or Noto Sans CJK SC
• for a formal academic look, keep title, abstract, and section headings in the same serious serif family instead of switching to decorative display fonts

Recommended visual combination for a polished Chinese translation:

• English text and math: newtxtext,newtxmath or another Times-like family
• Chinese body: Source Han Serif SC, Noto Serif CJK SC, or Songti SC
• Chinese headings: use the same serif family as the body by default
• heading fallback only when needed: Hiragino Sans GB
• monospace: Noto Sans Mono CJK SC
• title sizing: about 22pt to 26pt
• abstract sizing: around 10pt
• body leading: around 1.02 to 1.05
• avoid decorative handwriting, calligraphic, or UI-style Chinese fonts for paper titles and section headings unless the user explicitly wants a report-cover look

Recommended preamble patch template for translated copies:

\usepackage{newtxtext,newtxmath}
\usepackage[UTF8,fontset=none]{ctex}
% Prefer installed font family names first.
% If family lookup fails, replace these with explicit Path/Extension options
% that match fonts available on the current system.
\setCJKmainfont{Source Han Serif SC}
\setCJKsansfont{Source Han Serif SC}
\setCJKmonofont{Noto Sans Mono CJK SC}
\linespread{1.03}

Fontspec/XeCJK troubleshooting:

• If fc-list shows a font but XeLaTeX still says the font cannot be found, do not trust the family alias alone. Prefer Path=..., Extension=..., and the font basename.
• In fontspec option lists, every key-value pair must be comma-separated. Missing commas often surface as a misleading “Misplaced '=' in key-value input” error.
• If ctex auto-detects a problematic macOS fontset, switch to fontset=none and set CJK fonts explicitly.
• If Chinese headings look lighter than the body because no bold face is resolved, specify BoldFont or use AutoFakeBold for the CJK family.
• Keep one font strategy in the preamble. Do not scatter \setCJK... commands into section files.

If the document class or template already has fragile paragraph-opening effects, disable them in the translated copy when they break Chinese layout. Example: \IEEEPARstart often looks wrong for Chinese first paragraphs and may need to be replaced with a normal paragraph opening.

4. Split large jobs across subagents

When the paper is large, use multiple subagents in parallel.

Default threshold:

• more than 6 translatable .tex content files, or
• more than roughly 2,000 lines of prose-heavy LaTeX

Parallelization rules:

• keep one owner per file set to avoid merge conflicts
• split by independent section files whenever possible
• reserve the main agent for shared macros, preamble changes, global consistency, final review, and compilation
• do not assign bibliography, figures, or shared style files to translators unless they contain user-visible prose that must be localized
• do not let section translators insert font-selection commands into chapter files; keep font and preamble work in the main file unless the user explicitly asks otherwise

Suggested split:

• agent 1: abstract, introduction, conclusion
• agent 2: methodology sections
• agent 3: evaluation, applications, discussion sections

5. Translate only the right content

Apply the rules in references/translation-rules.md.

Minimum expectations:

• translate all user-visible English prose in title, abstract, headings, paragraphs, list items, and necessary captions
• keep formulas unchanged
• translate table and figure captions only as needed for reader comprehension
• translate only necessary comments inside code blocks or algorithm blocks
• do not translate person names
• do not translate bibliography entries unless the user explicitly asks for localized references

Prefer natural Chinese academic prose over literal word-for-word rendering, but do not add or remove substantive claims.

Do not redraw or OCR-translate text embedded inside figures or screenshots unless the user explicitly asks for image editing. If an image contains English labels, leave the image unchanged and translate only the surrounding caption when useful.

6. Review for missed English

After translation, review the _cn tree manually and with the helper script:

python3 scripts/check_translation_gaps.py /path/to/paper/<source-dir>_cn

Use the script output as a suspicion list, not as an oracle. Manually inspect any reported line before editing.

Review checklist:

• every included section file has been reviewed
• no section titles remain in English unless they are proper nouns
• no abstract or conclusion paragraphs remain in English
• captions are localized where needed
• no accidental edits broke macros, labels, or includes
• the translated directory still contains every required asset and bibliography file

7. Compile with XeLaTeX

Run:

python3 scripts/build_xelatex.py /path/to/paper/<source-dir>_cn

The script should auto-detect the main file, run the required XeLaTeX passes, and print the generated PDF path.

If the build fails:

• fix the translated copy, not the English source
• prefer minimal preamble fixes over sweeping reformatting
• rerun until the translated PDF builds cleanly

If the PDF renders but layout looks suspicious, inspect the output visually and, when helpful, use the pdf skill to rasterize or spot-check pages.

8. Final cleanup of the paper root

Before finishing, verify that the paper now lives entirely inside its dedicated subfolder.

Cleanup rules:

• keep the normalized paper folder
• keep shared directories such as docs/
• remove or move duplicate loose root-level copies of the same paper input when the normalized folder already contains the canonical English PDF and English source tree
• do not leave a second loose English PDF or a second loose source directory in paper/ for the same paper unless the user explicitly wants to keep the staging copies
• ignore harmless system files such as .DS_Store

Default expectation after cleanup:

paper/
  docs/
  <paper-slug>/

plus any other already organized paper folders.

Final Deliverable

Before finishing, provide:

• the paper folder path
• the translated source directory path
• the generated Chinese PDF path
• a note that the paper/ root was normalized and whether any loose intake files were removed or retained
• a brief note on any manual LaTeX compatibility fixes that were required

Do not claim completion until the translated PDF has been rebuilt successfully or you have clearly reported the blocking compile error.