New AI Tool Builds Paper Drafts and Verifies Citations Against Full Text

The tool breaks the process from topic selection to export into nine stages: determining the research scope, retrieving literature, cleaning and filtering, analyzing trends, finding an entry point, outlining, writing, verifying, and exporting. A user gives it a research topic, such as "RSI in the physical world," and the system extracts specific research questions, generates Chinese and English keywords, and searches OpenAlex, Crossref, Semantic Scholar, and arXiv, the report said.

Retrieved papers are deduplicated, scored, and filtered for relevance by a model. For highly relevant papers, the system also follows their references and later citations to add more literature, a strategy the report compares to Connected Papers. The collected papers are then organized into an evidence table that summarizes each paper's research question, method, data, conclusions, and limitations.

Using that table, Open Academic Paper Gen analyzes research trends and proposes possible research angles, gaps, and hypotheses. The report describes an "innovation diagnosis" step in which the system evaluates a proposed angle from the dimensions of problem, method, data, and perspective, identifies the closest existing work in the literature pool, and lists questions reviewers might raise. The report notes that a failure to find similar work in the retrieved pool does not mean no one anywhere has done it.

After the research angle is set, the system generates an outline and drafts the paper section by section. The draft then goes through citation verification and a simulated review before export in Markdown or LaTeX, with reference formats such as RIS and BibTeX.

The report says the citation check is split into three layers. First, every citation marker must match a paper that has already been collected; if the model invents a citation identifier, the system flags it. Second, the paper's identity is checked: papers with a DOI are matched against Crossref records and, if that fails, against doi.org. Mismatches in title, author, or year trigger warnings.

The third check moves from bibliographic information to content. When full text is available, the system finds the paragraphs that best match the cited claim and sends them, along with the beginning of the paper, to a model to judge whether the source supports the sentence in the draft. If full text is unavailable, it uses the abstract or an excerpt. When page data exists, warnings can include the corresponding PDF page for users to check.

According to the report, a lack of evidence found by the model does not prove that a claim is wrong. Open Academic Paper Gen can mark a claim as "unable to judge" when evidence is insufficient, and labels papers with too little text to verify as "unverified" rather than passing them. In its default mode, the workflow pauses at major stages so users can review, edit, and confirm before continuing; a fully automatic mode is also available.

The project supports OpenAI and Zhipu GLM models, but the report says claim verification, checking statements without citations, and simulated review still require an OpenAI API key. Open Academic Paper Gen comes from an individual developer named Maoshu, who has previously released projects related to AI screenwriting and AI research, according to the report.