论文精读翻译

ArXiv 论文精读级中文翻译，同步到 IMA 知识库 + 腾讯文档。 基于 5 篇论文（MDL/Kunlun/OneTrans/RankMixer/MixFormer）3 轮迭代实战经验。 触发场景：翻译论文、翻译 arxiv、论文精读、论文中文翻译、paper translation、 translate p...

Security Scan

The skill's stated purpose (paper translation + uploads) matches its instructions, but it omits declaring required CLI tools and the credentials/config needed to perform uploads — this mismatch is a notable incoherence you should understand before installing.

Purpose & Capability

The skill's name and description (ArXiv paper translation and upload to IMA + Tencent 文档) match the runtime instructions: downloading arXiv HTML/images, translating, validating, and calling upload APIs. However, the skill does not declare any required credentials or configuration even though the SKILL.md shows it will call mcporter to upload to Tencent and IMA; this omission is unexpected.

Instruction Scope

SKILL.md stays on task: fetch arXiv HTML once, download images, perform paragraph-level translation, run the included validation script, and upload two Markdown variants. It does include steps for interactive auth (generate code, poll for token) and calls to external tooling (curl, jq, mcporter, python3). It does not instruct reading arbitrary unrelated local files or exfiltrating environment variables, but it will transmit the translation and images to external services as intended.

Install Mechanism

No install spec (instruction-only) and a small Python validation script — low install risk. But the flow relies on external CLIs (mcporter, jq, curl, web_fetch, python3). Those tools are not provided by the skill and are required at runtime; the skill does not document them in the declared requirements.

Credentials

The skill declares no required environment variables or credentials, yet SKILL.md describes authenticating with Tencent/IMA via mcporter and obtaining tokens. Uploading to cloud services will require credentials/config; not declaring them is an inconsistency and hides what secrets the skill will need/access at runtime.

Persistence & Privilege

always is false and the skill does not request persistent/automatic always-on presence. It does not claim to modify other skills or global agent settings. Autonomous invocation is allowed by default (normal).

What to consider before installing

Before installing, verify these points: - Expect to need external CLIs: mcporter, curl, jq, python3 (and whatever web_fetch is). The skill's metadata lists none of these — confirm they exist in your environment. - The SKILL.md uploads documents/images to IMA and Tencent Docs via mcporter. That will require service tokens/credentials; ask the skill author which env vars/config files are required and where credentials are stored. - Inspect/verify the mcporter CLI and any endpoints it calls (mcporter is a third-party tool — ensure it is trusted and review its auth/storage behavior). - The included validate_translation.py is benign and only reads the markdown file to check formatting; it does not exfiltrate secrets. - If you will provide tokens, follow least-privilege practices (use an account with minimal rights) and avoid giving org-level write access until you test with a non-sensitive example. - If you need stricter guarantees, request the author to declare required binaries and environment variables in the metadata, and to document exactly how tokens are obtained/used and where uploads are sent.

ArXiv 论文精读翻译

Base directory for this skill: {SKILL_DIR}

将 ArXiv 论文逐段翻译成中文，生成双版本 Markdown（IMA + 腾讯文档），并上传到两个平台。

三条铁律

1. 完整翻译不精简 — 逐段翻译每个 paragraph，不遗漏任何论证细节。大模型倾向于"帮你归纳"，但用户要的是精读级翻译。
2. 译注显式标记 — 大模型解读必须用 > **[译注]**：... 引用块，绝不混入原文翻译。
3. 简称首次标全称，后续直接用 — 首次出现标注全称并核对原文，后续不再展开。避免错误展开（如 TA=Target Attention 被误写为 Transformer Aggregator）。

标准 6 步流程

Step 1: 获取原文

web_fetch https://arxiv.org/html/<id>v<n>

• 只 fetch 一次，节省 token
• 同步下载图片：curl -sL -o x{n}.png https://arxiv.org/html/<paper_id>/x{n}.png
• 下载后检查文件大小，相同大小的异常文件（404 垃圾响应）删除

Step 2: 翻译生成

• 逐段翻译，不做精简
• 首行元信息：原标题、arxiv 链接、年月、机构、翻译辅助大模型名称
• 简称首次出现标全称（核对原文），后续用简称
• 译注用 > **[译注]**：... 格式
• 结构化排版：多级标题 + 列表 + 加粗 + 表格 + 引用块
• 公式保留 LaTeX；\bm 全部替换为 \boldsymbol
• 图表按原文顺序插入所在章节标题之后、小节正文之前
• 参考文献完整列出

表格处理策略：

• 简单表格 → Markdown 表格重写（可搜索/编辑）
• 复杂表格（合并单元格/特殊排版）→ PyMuPDF 从 PDF 截取

Step 3: 自动化校验

翻译完成后，运行校验脚本：

python3 {SKILL_DIR}/scripts/validate_translation.py <markdown_file>

校验项：

检查项	标准
章节完整性	包含：摘要/引言/相关工作/方法/实验/结论/参考文献
LaTeX 兼容性	\bm 出现次数 = 0
译注标记	数量 > 0，格式为 > **[译注]**
参考文献	条数列出供人工核对
图片链接	外链格式正确

Step 4: 生成两版 Markdown

• IMA 版：图片用 arxiv 外链 URL / base64 data URI
• 腾讯文档版：用脚本从 IMA 版自动替换图片链接为 image_id

图片上传流程：

1. curl -sL -o x{n}.png https://arxiv.org/html/<paper_id>/x{n}.png 下载
2. 腾讯文档：mcporter call tencent-docs upload_image → 拿 image_id
3. IMA：直接用 arxiv 外链 URL

详见 references/platform-compat.md

Step 5: 上传 IMA 知识库

# create_media → COS 上传 → add_knowledge（media_type=7 = Markdown）
# 如遇 code=220030（限流），sleep 15s 重试，cos_key 仍有效

Step 6: 上传腾讯文档

TITLE="【YYYY.MM｜组织】XXX 中文翻译"  # 必须 ≤36 字符
jq -n --arg title "$TITLE" --rawfile mdx "$FILE" --arg cf "markdown" \
  '{title:$title, mdx:$mdx, content_format:$cf}' > /tmp/args.json
mcporter call tencent-docs create_smartcanvas_by_mdx --args "$(cat /tmp/args.json)"

mcporter 传大参数不支持 --args-file，必须用 --args "$(cat file.json)"。

命名规范（强制）

平台	格式	约束
腾讯文档标题	【YYYY.MM｜组织】XXX 中文翻译	≤36 字符（按字符数，非字节）
IMA 文件名	【YYYY.MM｜组织】XXX 中文翻译.md	同名加 .md

• 两平台必须完全一致，不加 v2/图文版 等后缀
• 示例：【2026.02｜ByteDance】MixFormer 中文翻译

翻译后 Checklist

完成翻译后逐项确认：

• grep -c '\\bm' = 0
• 简称首次出现已标全称且正确
• 译注均用 > **[译注]**：... 格式
• 图表位置与原文章节顺序一致
• 参考文献条数与原文一致
• IMA 版和腾讯文档版图片格式各自正确
• 两平台文件名/标题完全一致
• 首行包含论文元信息（标题/链接/年月/机构/大模型名称）

效率优化

• 只 web_fetch 一次原文（节省 token）
• 直接生成最终版，不生成中间草稿（减少 50%+ 工具调用）
• 图片下载 + 上传并行执行
• 用脚本自动从 IMA 版生成腾讯文档版
• 自动化校验脚本在上传前拦截格式问题

参考文档

• 踩坑经验 + 平台兼容性：references/platform-compat.md — LaTeX 兼容、图片跨平台、API 限流等详细说明
• 迭代历史：references/iteration-history.md — MixFormer v1→v3 的完整教训记录