New open-source tool Open Academic Paper Gen drafts papers and checks each citation against the source text

A new open-source project called Open Academic Paper Gen has landed on GitHub, according to Chinese tech outlet QbitAI. The tool not only produces a draft complete with references; when it can access the full text of a cited paper, it goes back into the source to locate the relevant passage and check whether that paper actually supports the claim attached to it.

That verification angle is the project's selling point. QbitAI notes that AI-generated claims and real papers can coexist happily until a claim is put back into the full text and read in context, at which point it turns out the original author may have meant something else entirely. To address this, Open Academic Paper Gen locates the passage that best matches a cited claim and judges whether the paper can support the statement, annotating the corresponding PDF page when page data is available.

Structurally, it is a multi-agent workflow for generating paper drafts, splitting the process from topic selection to export into nine stages. Given a research topic, it derives concrete research questions, generates Chinese and English keywords, and searches for papers across OpenAlex, Crossref, Semantic Scholar and arXiv.

Results are deduplicated and scored, then filtered for relevance by a model. For highly relevant papers, it also follows their references and the later work that cites them, snowballing outward. The collected literature is compiled into an evidence table that summarizes each paper's research question, method, data, conclusions and limitations.

From there the system analyzes research trends and proposes research angles, gaps and hypotheses, including a novelty check that evaluates a candidate direction across problem, method, data and perspective, flags the closest existing work in the pool, and anticipates objections a reviewer might raise. Once an angle is fixed, it generates an outline and drafts the chapters.

After the draft, citations are verified and a simulated peer review runs. The tool can export Markdown or LaTeX, plus reference formats such as RIS and BibTeX. Citation checking is split into three layers: reference markers must map to collected papers, with invented citation keys flagged; papers with a DOI are checked against Crossref and then doi.org, with title, author or year mismatches triggering warnings; and at the content level the system extracts the passage that best matches a cited claim and asks a model whether the original text supports it.

Where evidence is thin, the model is allowed to return a cannot-determine label, and papers with too little text to check are marked unverified rather than passed. By default the workflow pauses at major stages so a user can review, edit and confirm before continuing, though a fully automatic mode is available.

The project comes from an individual developer, mmlong818, who has previously released AI screenwriting and AI research projects. One practical caveat is that while it supports OpenAI and Zhipu GLM, the argument-verification, uncited-statement check and simulated-review steps still require an OpenAI API key.