Researchers in China, Japan and Korea run whole papers through coding agents and post how

The sources are public posts and repositories in three languages. The translations are ours, and each claim is the poster’s own.

China packages the whole paper as skills

QbitAI reported the first project at 6,400 stars in May. GitHub showed 51,081 on 9 October. Its README, as QbitAI quotes it, says “AI is your copilot, not the pilot” (AI是你的副驾驶，不是飞行员).

The Zhihu post is blunt about the reader it is written for: anyone still pasting paragraphs into ChatGPT to polish them is “two full technology generations behind” (落后了整整两个技术代差). It comes from a commercial account that sells research services, so read its timing claim as a sales claim.

Japan publishes the logbook

A data scientist at the Matsuo Institute wrote on Zenn in July 2026 about a natural language processing paper submitted to ACL Rolling Review. Claude Code acted as conductor (指揮者), handing experiment design to Codex CLI and paper reading to Gemini CLI.

The author set two rules: every number in the text is checked against the raw result files, and every citation is checked by hand against the original. The lesson the author draws is that agents made the loop of training, evaluating and analyzing far faster, and that deciding what the results mean stayed with the person.

A master’s student on Zenn describes a simpler setup: a CLAUDE.md holding academic style rules, LaTeX rules, a term list and the thesis outline. The agent drafts each chapter and the student adjusts it. The post’s warning is that “AI takes no responsibility” (AIは責任を持たない).

Korea reports the jump in one year

On the graduate school board of DC Inside, a poster wrote on 1 May 2026 about having Claude Code build variants of an open source simulation code. The result was “far better” than the same attempt a year earlier, and the code came out “clean, at the level I would write myself” (내가 직접 코딩한 수준으로 깔끔하게). The poster adds that the agent was poor at finding the cause of its own small errors and needed detailed instructions for each fix.

Korean brokers are further along than Korean campuses. Mirae Asset moved research assistant work to AI and promoted the assistants.

Nobody here hands over the checking

None of the posts concerns finance, and none reports a paper accepted at a journal. They show the method, and the method carries over: an agent as conductor, the paper as a set of skills, and a gate on citations and numbers before anything is written up. The pinned SEC guide applies the same pattern to a finance dataset.