Polish润色
Rewrite under the writing rules, or under your own instruction. Claims, numbers, citations and formulas stay fixed.
按写作规则改写，也可以写上你的要求。论断、数字、引用和公式保持不变。
One draft.
Two languages.
一篇稿子，两种语言。
一篇稿子
两种语言。
One draft. Two languages.
Write an academic paper in English and Chinese at once. Type on one side; the other keeps up paragraph by paragraph. LaTeX compiles beside the source, and the writing assistant follows the rules of research-builder.
在一个窗口里同时写中英文论文。在一侧写作，另一侧逐段跟上。LaTeX 在源码旁编译，写作助手遵循 research-builder 的规则。
Free download · no account · bring your own model
免费下载 · 无需账号 · 使用你自己的模型接口
Why write in two languages为什么用两种语言写作
The premise 出发点
A translation is not an attachment you bolt on at the end. It is the second reading of your paper — the one many of your readers will actually use.
译文不是最后才补上的附件，
而是论文的第二种读法——
读者真正会去读的那一种。
BiWrite keeps both readings alive while you write.
BiWrite 让两种语言在你写作的时候同时成立。
-
01
Open a paper
打开论文
A
.tex,.mdor.txtfile. If you already keep a Chinese version, pair it and BiWrite uses it.支持
.tex、.md、.txt。已经有中文版就配对进来，BiWrite 直接使用你的译文。 -
02
Write once
只写一遍
One side is yours to edit. The other follows paragraph by paragraph, revising only what changed.
一侧由你编辑，另一侧逐段跟随；只修订改动过的段落，其余一字不动。
-
03
Finish in both
双语定稿
Compile the PDF beside you, polish with rules, and export the Chinese version when you need it.
PDF 在旁边编译，按规则润色，需要时导出中文版本。
01
Bilingual editing双语编辑
The other language keeps up with you.
另一种语言，
跟着你的笔走。
English on the left, Chinese on the right — or the other way round. Stop typing for a moment and the paragraph you changed is translated again, revised from its previous translation. Nothing else is touched.
左侧英文，右侧中文，也可以反过来。停下打字，改过的段落就会以原译文为基础重新修订；其余段落，一字不动。
-
Only edited paragraphs. A changed paragraph is revised from its old translation, not translated from scratch, and the result streams in.
只重译改过的段落。以原译文为基础修订，而不是推倒重来；结果以流式返回。
-
Placeholders, both ways. Math, citations, cross-references, labels, URLs and comments reach the model as placeholders and return byte for byte.
占位符来回无损。公式、引用、交叉引用、标签、网址和注释以占位符发给模型，返回后逐字节还原。
-
Switch or pair.
EN ↔ 中puts Chinese on the left to edit. Pairpaper.texwithpaper_zh.texand your own Chinese is used — with no translation requests.换边或配对。用
EN ↔ 中把中文放到左侧编辑。把paper.tex与paper_zh.tex配对，你的中文直接作为译文——不产生翻译请求。
paper.tex
15 translated · 6 skipped已翻译 15 · 跳过 6
EnglishEnglishpaper.tex
7
Graph neural networks (GNNs) are the standard tool for learning on relational data [kipf2017gcn]. Their success rests on task-specific training.
8
We answer this question affirmatively. Our contributions are threefold. First, we formalize in-context learning for graphs. Second, we propose a prompt graph construction that connects query nodes to labelled example nodes.
9
\section{Method} Let $X \in \mathbb{R}^{n \times d}$ denote node features and $A$ the adjacency matrix.
中文译文
7
图神经网络（GNN）是用于关系数据学习的标准工具 [kipf2017gcn]。其成功依赖于针对特定任务的训练。
8
我们对这一问题给出了肯定的回答。我们的贡献有三个方面。首先，我们对图上的上下文学习进行了形式化定义。其次，我们提出了一种提示图构建方法，将查询节点与带标签的示例节点连接起来。
9
\section{方法} 令 $X \in \mathbb{R}^{n \times d}$ 表示节点特征，$A$ 表示邻接矩阵。
¶ 8 edited · revised from previous translation
¶ 8 有改动 · 在原译文基础上修订
1 request · 2.4 s
1 次请求 · 2.4 秒
Placeholders, both ways占位符，一来一回
sent
… formalize in-context learning for ⟦1⟧. … match the baseline on ⟦2⟧.
back
⟦1⟧ $G = (V, E)$ ⟦2⟧ \cite{kfp2017gcn}
Restored byte for byte. The model never rewrites your formulas or citation keys.
逐字节还原。模型不会改写你的公式或引用键。
02
LaTeX, in placeLaTeX，原地编译
The PDF sits beside the source.
PDF 就在
稿纸旁边。
Compile with your own TeX distribution. Unsaved edits compile in a hidden cache, so the file on disk stays exactly as you last saved it. Click a sentence in the PDF and its source is selected.
用你本机的 TeX 发行版编译。未保存的修改在缓存里编译，磁盘上的 .tex 保持你上次保存的样子。点 PDF 里的句子，源码中的那一句随即被选中。
-
Unsaved, still compiled. Edits are written to a
.biwrite/cache folder; your paper on disk does not move.不保存也能编译。修改写入
.biwrite/缓存文件夹，磁盘上的论文保持原样。 -
Click to source. A click on a PDF sentence selects it in the source through SyncTeX.
⌘⌥Jfinds the cursor's paragraph in the PDF.点 PDF，回到源码。由 SyncTeX 定位到对应句子；
⌘⌥J在 PDF 中高亮光标所在的段落。 -
A Chinese PDF too. XeLaTeX and
ctexbuild a PDF from the translation; the Chinese.texcan be exported on its own.还有中文 PDF。用 XeLaTeX 和
ctex由译文编译，也可以单独导出中文.tex。 -
Any engine, real templates. latexmk, pdfLaTeX, XeLaTeX or LuaLaTeX. PVLDB 2027, ICLR 2027 and IEEE TII are built in.
各种引擎，真实模板。latexmk、pdfLaTeX、XeLaTeX、LuaLaTeX。内置 PVLDB 2027、ICLR 2027 和 IEEE TII。
03
Writing assistant写作助手
A second reader who knows the rules.
多一位懂规矩的读者。
Select a sentence and ask. Polish it, question it, write the next paragraph in the manner of your references, draw the figure, or repair the LaTeX. Jobs run in the background and results arrive as diffs; nothing enters the paper until you accept it.
选中一句，然后发问。润色、提问、按参照样例仿写下一段、画图、修 LaTeX。任务在后台运行，结果以差异显示；你点采纳，它才进入论文。
In a LaTeX paper, Accept first has the model fit the result to your document, then compiles it. If it needs a package or a TikZ library, the model repairs that before the text is written.
在 LaTeX 论文中，采纳会先让模型把结果改成适合这篇文档的样子，再编译一次。缺宏包或缺 TikZ 库时，模型会先修好，再写入正文。
Wait as long as you like. Edits you make while a job runs are not overwritten — the result waits for you.
可以慢慢等。任务运行期间你继续写的内容不会被覆盖，结果会放在那里等你。
Ask提问
Questions about the selection or the whole paper, with follow-ups. Answers, never edits.
就所选文字或整篇论文回答，可以追问。只给回答，不改正文。
Imitate仿写
A new paragraph after the cursor, in the voice of your writing samples, with a facing Chinese version.
在光标后写新段落，模仿你写作样例的文风，并附对照中文。
Figures图表
LaTeX code, a compiled PDF, or an editable PowerPoint diagram — traced from a draft or only in its style.
LaTeX 代码、编译好的 PDF，或可编辑的 PowerPoint 示意图——可以临摹，也可以只仿风格。
Fix语法修复
TeX errors handed to the model with their lines, up to three rounds, shown as a diff before anything changes.
把 TeX 错误连同所在行交给模型，最多三轮；改动以差异显示，你确认后才写入。
Via translation按译文改
Rewrite the Chinese; the English follows with the smallest edit that means the same thing.
改写中文，英文以最小的改动成为同一个意思。
The method behind the assistant助手背后的方法
Not another chatbot in the sidebar.
不是又一个
塞进侧栏的聊天框。
Every assistant action follows a part of research-builder — an open, MIT-licensed method for empirical papers. BiWrite ships a copy of the skill, can update it from GitHub, and can use your own skill folder together with its materials/.
每一项助手操作都遵循 research-builder 的一部分。那是一套开源（MIT）的实证论文方法。BiWrite 自带一份 skill，可以从 GitHub 更新，也可以换成你自己的 skill 文件夹，连同其中的 materials/ 一起使用。
01
De-AI phrasing去 AI 腔
Remove the tells. Keep the argument.
去掉机器味，留住论证。
02
Voice matching声音匹配
Write in the voice of your own samples.
用你自己样例里的声音来写。
03
Paper anatomy论文骨架
New paragraphs follow the paper's structure.
新段落遵循论文本来的结构。
04
Figure style图表风格
Figures in the method's visual language.
图遵循这套方法的视觉语言。
05
Editable PPT diagrams可编辑 PPT 图
Native shapes, arrows and your icons.
原生形状、块状箭头和你的图标。
06
Your materials你的素材
Your papers, figures and icons as references.
你的论文、图和图标可以作为参照。
04
Also in the box箱子里还有
Small things that save the paper.
那些真正省事的细节。
Templates模板
PVLDB 2027, ICLR 2027 and IEEE TII are built in. Import a folder or a .zip as a template; export a project as one; open any LaTeX folder.
内置 PVLDB 2027、ICLR 2027 和 IEEE TII。把文件夹或 .zip 导入为模板，把项目导出为模板，也可以打开任意 LaTeX 文件夹。
PVLDB · ICLR · IEEE TII
Your own Chinese version你自己的中文版
Pair paper.tex with paper_zh.tex, or sections_en/ with sections_zh/. Your Chinese is the translation, with no requests, and every edit is written back in place.
paper.tex 与 paper_zh.tex，或 sections_en/ 与 sections_zh/ 逐段配对。你的中文直接作为译文，不产生请求；改动原位写回另一个文件。
21 / 21 paragraphs paired
Any model provider接上任何模型
OpenAI-compatible Chat Completions, the OpenAI Responses API and Anthropic Messages, with presets for DeepSeek, Qwen, Kimi, OpenRouter and local Ollama.
OpenAI 兼容的 Chat Completions、OpenAI Responses 和 Anthropic Messages；DeepSeek、通义千问、Kimi、OpenRouter、本地 Ollama 等都有预设。
Chat · Responses · Messages
Key pools号池
Several named keys per provider, used evenly with a per-key limit, failover on rate limits, and a new key swapped in place. Import and export to a file.
每个接口挂多个命名的密钥，按并发上限均匀使用，限流时换号；可以原位换 key，也可以导入和导出到文件。
keys × concurrency
Request log请求日志
The model, reasoning effort and service tier that were sent — and the ones the server declared. Timings, tokens and the key used, without ever storing the key.
记录发出的模型、推理强度和服务等级，以及服务端声明的实际值；还有耗时、token 和所用密钥，但从不保存密钥本身。
1.2k in · 300 out · 2 cached
My materials我的素材
writing-samples/ sets the voice; domain/ and figures/ are references; icons/ go into PowerPoint figures. The folders mirror the skill's materials/.
writing-samples/ 给出你的文风；domain/ 与 figures/ 作为参照；icons/ 用在 PPT 图里。分类与 skill 的 materials/ 相同。
voice · references · icons
Any GitHub release can be installed from inside the app, pre-releases and older versions included. Closing the window leaves BiWrite in the menu bar or notification area, so the paper is still open when you come back. 软件内可以安装 GitHub 上发布的任意版本，包括预发布的分支构建和旧版本。关闭窗口后，BiWrite 带着打开的文档留在菜单栏或任务栏通知区域，回来时论文还在。
Get BiWrite下载 BiWrite
Write it once.
写一遍就够了。
Download the latest release for macOS or Windows. Bring a model endpoint, and the TeX distribution you already use.
下载适用于 macOS 或 Windows 的最新版本。准备一个模型接口，以及你本来就在用的 TeX 发行版。
Version 0.2.25 · free download版本 0.2.25 · 免费下载
You'll need你还需要
- macOS 12+, or Windows 10 / 11.
- macOS 12+，或 Windows 10 / 11。
- A TeX distribution for LaTeX: MacTeX, TeX Live or MiKTeX.
- LaTeX 功能需要 TeX 发行版：MacTeX、TeX Live 或 MiKTeX。
- LibreOffice, only to turn PowerPoint figures into PDF.
- 把 PPT 图转成 PDF 时需要 LibreOffice。
- A model endpoint for translation and the assistant.
- 翻译和写作助手需要一个模型接口。
First launch首次打开
The installers are not signed with a trusted certificate yet. On macOS, open BiWrite once, then allow it under System Settings → Privacy & Security → Open Anyway. On Windows, SmartScreen: More info → Run anyway.
安装包尚未获得受信任的证书签名。macOS 上先打开一次，再到系统设置 → 隐私与安全性点「仍要打开」；Windows 上在 SmartScreen 中依次点「更多信息」「仍要运行」。
Your data你的数据
Keys live in the system keychain. Settings, the translation cache and request logs stay on your machine. No account, no analytics.
密钥保存在系统钥匙串。设置、翻译缓存和请求日志只存在本机。没有账号，也没有统计。
Is BiWrite free?BiWrite 免费吗？
Yes. Download the installers from GitHub Releases at no cost. You pay your model provider directly, if it charges anything.
免费。安装包在 GitHub Releases 上直接下载。模型接口的费用由你直接付给服务商（如果它收费）。
Does it work offline?可以离线使用吗？
Opening, writing, compiling and paired translations work offline. New translations and assistant jobs need your model endpoint, which can be a local server such as Ollama.
打开、写作、编译和已有的配对译文都可以离线完成。新的翻译和助手任务需要模型接口，也可以接本地的 Ollama。
What happens to my files and keys?文件和密钥会被怎样处理？
Your files stay where they are; edits are only written when you save. Keys go into the system keychain and are never shown again. Request logs stay on this machine and do not record keys.
文件留在原来的位置，只有保存时才写入。密钥保存在系统钥匙串中，之后不再显示。请求日志只存在本机，不记录密钥。
Can I keep my existing Chinese version?可以继续用我已有的中文版吗？
Yes. Pair paper.tex with paper_zh.tex, or sections_en/ with sections_zh/. BiWrite matches paragraphs by citations, labels, formulas and numbers, and writes each edited paragraph back in place.
可以。把 paper.tex 与 paper_zh.tex 配对，或 sections_en/ 与 sections_zh/。BiWrite 按引用、标签、公式和数字匹配段落，保存时把每个改动过的段落原位写回。