Glonce

Researchers propose Agent-Native Research Artifact to replace scientific papers

researchThe Glonce editorsIEEE Spectrum🔥75Glonce rating75Impact64Novelty72Relevance98Surprise67Players59Credibility83
Researchers propose Agent-Native Research Artifact to replace scientific papers

In May 2026, 37 researchers from roughly two dozen top universities and tech companies published a paper on arXiv titled "The Last Human-Written Paper," arguing that scientists should stop writing papers for human readers. The authors write that AI agents are becoming first-class participants in research workflows, not tools that assist humans but autonomous contributors that read, reproduce and extend scientific work, and that this transition demands infrastructure built around agents from the start. Their proposed replacement is the "Agent-Native Research Artifact" (ARA), a format built for AI agents to use efficiently; the paper itself is published online in ARA form as a demonstration.

IEEE Spectrum interviewed lead author Jiachen Liu about the proposal. Liu conducted the work on ARA while completing her Ph.D. in computer science at the University of Michigan, awarded in 2025, and this May cofounded the Agent Native Research Lab, an AI-for-science startup in Palo Alto, California. She traces her change of view to the end of 2024, when the Cursor coding agent appeared: she initially thought AI still needed extensive human-built "harness" infrastructure to be useful, but says models have advanced enough since then that most Ph.D.-level and professor-level knowledge will soon reside inside them, at which point humans stop being able to add value and AI needs infrastructure to evolve on its own.

Liu and her co-authors identify two flaws in the traditional paper format from an AI's perspective. The first is the "storytelling tax": she says 80 percent of the information generated during a piece of research is lost once it is written up, because only the final 20 percent, the polished narrative, survives, while the failed attempts, side branches and small decisions that actually made the work succeed are discarded. The second is the "engineering tax": even the 20 percent that does survive is a lossy compression, often too ambiguous or missing implementation details for another party to reproduce.

To address this, the ARA proposal includes a component called the "Live Research Manager," an AI system that observes and documents a researcher's work automatically as it happens, without requiring the researcher to write anything down; a paper in the traditional PDF format can then be generated from that record when needed. On the question of checking AI-generated results for errors and hallucinations at scale, Liu says a human cannot manually verify everything an AI produces, so the answer is a formal system that judges results objectively, avoiding the use of one language model to police another's work. She says she is separately developing a system using neurosymbolic techniques, combining neural networks with logic-based structures, so that every claim could in principle be expressed as a formula and proved, though the article gives no further detail, timeline or evidence that this system currently works.

Asked what happens to researchers' reluctance to expose mistakes and dead ends, Liu argues that AI removes the stigma: a human overseeing an AI that spent 12 unproductive hours can point out its mistake without personal embarrassment. She also referenced two other essays of hers, "The End of Human-in-the-Loop" and an earlier piece emphasizing the human's importance in the loop, saying she now expects a point where AI, having absorbed all available expert knowledge from humans, will no longer need human input and will begin to self-evolve. She said she believes junior scientists will still develop the necessary skills, just through a different learning process built around working with AI. The article notes the paper's authors are not named beyond Liu, nor are their specific institutions listed, and it gives no information on the paper's peer-review status, adoption by any journal or institution, or citation activity since publication.

Key facts

  • 37 researchers from roughly two dozen top universities and tech companies published "The Last Human-Written Paper" on arXiv in May 2026, arguing papers should be written for AI agents rather than humans.
  • The paper proposes the Agent-Native Research Artifact (ARA), a format built for AI agents; the paper itself is published online in ARA form.
  • Lead author Jiachen Liu, who earned her Ph.D. from the University of Michigan in 2025, cofounded the Agent Native Research Lab in Palo Alto this May.
  • Liu says the traditional paper format loses 80 percent of a research process's information (the "storytelling tax") and that even the surviving 20 percent is often too ambiguous to reproduce (the "engineering tax").
  • ARA includes a "Live Research Manager" that automatically observes and documents research as it happens, and Liu is separately developing a neurosymbolic system meant to let AI-generated claims be formally proved rather than checked by another language model.

Why it matters

The proposal reframes AI not as a tool that assists researchers but as an autonomous participant in the research process, and argues the entire infrastructure of science, the paper itself, should be redesigned around AI as the primary reader. That is a significant claim about where AI capability is heading in scientific work, made not by a lab's marketing but by 37 researchers across roughly two dozen institutions in a formal proposal.

Who it affects

Researchers and institutions that produce and read scientific papers, journals and conferences that would need to support a new artifact format, and AI-for-science startups such as Liu's Agent Native Research Lab that are building infrastructure around this idea. The article does not say any journal, conference or institution has adopted or endorsed ARA.

How to use it

The ARA paper itself is published online in ARA form as a working example of the format, referenced in the article as available alongside the traditional arXiv posting. No pricing, licensing or product is described; the Live Research Manager and the neurosymbolic verification system Liu mentions are described as work in progress, without a release timeline.

How solid is it

The claim comes from 37 co-authors across about two dozen universities and companies, which is a broad base for a single proposal, and the paper is published on arXiv, a preprint server that does not itself imply peer review. The article does not state the paper's peer-review status, does not name the other 36 authors or their institutions, and cites no adoption, citation or reaction data beyond Liu's own description of "diverse... all positive" feedback.

Risks and caveats

The central mechanism for catching AI errors at scale, a neurosymbolic system that would let claims be formally proved, is described only in broad terms by Liu, with no timeline, implementation detail or evidence it currently works. Liu's own framing, that humans will eventually be reduced to a bottleneck AI no longer needs, is a specific and contestable prediction rather than an established fact, and the piece is an interview built around one author's account of a paper she led, not an independent assessment of the proposal's prospects.

“One is the "storytelling tax." Once we write everything into a paper, 80 percent of the information about the work is lost. We only write down the last 20 percent.”

— Jiachen Liu, lead author of "The Last Human-Written Paper"