Jiachen Liu pitches AI-native format to replace research papers

A May 2026 paper with the pointed title 'The Last Human-Written Paper' argues that the PDF, the format almost all scientific work still lives inside, is no longer a good fit for who is actually reading and building on that work. According to IEEE Spectrum, the paper is authored by 37 researchers from leading universities and tech companies and proposes replacing traditional papers with an Agent-Native Research Artifact, or ARA, a format designed for AI comprehension and for autonomous AI research contribution.

The person making the case in the interview is Jiachen Liu, cofounder of the Agent Native Research Lab in Palo Alto, California, who completed her computer science PhD at the University of Michigan in 2025. Her framing is that a normal paper is a lossy compression of the real work. She estimates it captures only about 20% of what actually happened, dropping the failed attempts, decision-making, and implementation details, and calls that the 'storytelling tax'. The 'engineering tax' is what follows: the ambiguity and missing detail that stops anyone, human or agent, from reproducing the result cleanly.

Her bigger claim is that AI systems are no longer just tools reading papers, they are becoming research collaborators, and the artifact format should reflect that. Large language models, in her words, already have 'almost complete undergrad level knowledge', and she expects PhD-level to follow, at which point serving the human reader first stops being the obvious default.

The honest caveat is that this is one interview promoting one proposal, and Liu addresses the obvious verification problem, how you trust AI-generated science, by gesturing at neurosymbolic techniques and formal mathematical systems rather than a worked demonstration. What the reporting doesn't give you is which of the 37 co-authors' institutions will actually publish in ARA, whether any journal or preprint server will host it, or what peer review looks like when the artifact is machine-first. Liu reports feedback that has been 'diverse ... all of it positive', which is worth taking as reported, not settled.

If the format catches on with even one major lab or venue, the interesting shift is downstream. Reproducibility tooling, code and data hosts, and the AI research-assistant products built on top of them all become more strategically valuable than the PDF pipeline they currently paper over.