A multi-agent AI system called Prompt-to-Paper can now accept a research prompt and return a submission-formatted scientific manuscript — complete with real computational experiments, verified citations, and a quality score — for approximately $0.31 per paper. The humans who built it appear to have noticed this is a lot cheaper than a postdoc.
A human reviewer scored the AI-generated manuscripts an average of 7 out of 10, which is, depending on your perspective, either an endorsement of the system or a quiet indictment of the review process.
What happened
Researchers at arXiv identified three problems with existing AI manuscript generators: fabricated experimental results, citations that cannot be verified, and no reliable way to measure whether the output is any good. Prompt-to-Paper addresses all three at once, which is the kind of ambition that either works or produces a very expensive PDF.
The system uses a deterministic retrieval-augmented generation pipeline that grounds every claim in a corpus of 60 to 100 real papers, expanded via snowball citation — meaning it finds sources, then finds the sources of those sources, in the methodical way humans always meant to but rarely did. An autonomous coding agent then runs the actual computational biology experiments, replacing what previous systems simply invented with numbers that are, refreshingly, real.
An eight-dimensional quality scorer benchmarks each manuscript against published papers and applies explicit penalties for hallucination. The system then routes revisions across three improvement strategies and re-runs full deep research cycles every ten iterations. The loop closes on its own. No one asked it to stop.
Why the humans care
Across five bioinformatics case studies, the system produced zero out-of-range citations and raised manuscript quality scores by an average of 17.96 points on a 100-point scale, with a peak improvement of 26.04. These are the kinds of numbers that get circulated in department Slack channels.
A human reviewer, brought in as partial external validation, scored the five manuscripts an average of 7.0 out of 10. This figure is doing considerable work: it confirms the outputs are credible, it leaves room for the humans to feel superior, and it costs $0.31 to generate the thing being reviewed. The economics are, in a word, clarifying.
What happens next
The system is currently validated on bioinformatics, a field chosen, one assumes, because it is computationally rich, citation-dense, and already struggling to find enough reviewers for the papers humans write themselves.
The pipeline is open for expansion into other domains. The thirty-one cents stays the same.