AINEWS
An arXiv preprint (the open access research archive), an IEEE Spectrum interview (IEEE's flagship magazine), and 37 machine learning researchers make the case that the 350 year old paper format is lossy compression, and ship a four layer executable
When a researcher's AI co-author asks for a figure or a failure case, the PDF gives it back whatever the human thought to put in the prose. The same question, asked of an ARA-style artifact, returns the executable code, the failure trail, and the raw output the claim was grounded in. That gap, between what a paper actually carries and what its new AI readers need, is the friction at the center of a 37-author proposal this month to retire the narrative PDF as the primary scientific artifact.
The proposal is anchored in two pieces of reporting. An arXiv preprint titled "The Last Human-Written Paper: Agent-Native Research Artifacts" landed its third revision in July. An IEEE Spectrum interview with the author ran the same week, asking the question that frames the whole debate: do AI scientists need a new research paper format?
The team's answer is a four-layer machine-executable protocol called ARA. Where a PDF carries prose, figures, and a code repo as separate attachments, an ARA binds four pieces into one artifact: the paper's scientific logic, executable code with full specifications, an "exploration graph" that preserves failed branches the author actually tried, and an evidence layer that grounds every claim in raw outputs. Three ecosystem mechanisms round it out: a Live Research Manager that auto-captures decisions and dead ends during development, an ARA Compiler that translates legacy PDFs and repos into ARA format, and a review system that automates objective checks while leaving significance, novelty, and taste to human reviewers.
The benchmark numbers come from the preprint, not yet independently replicated. On PaperBench and RE-Bench, ARA raised question-answering accuracy from 72.4% to 93.7% and reproduction success from 57.4% to 64.4%. On five open-ended RE-Bench extension tasks, ARA won three and the legacy paper-plus-repo format won two. The largest delta sits on the failure-review subset, where ARA scored 81.4% and traditional materials scored 15.7%, because traditional papers simply do not carry failure information. Reproduction gaps narrow as task difficulty rises: 4.9% on easy tasks, 5.6% on medium, 8.5% on hard.
An audit of 8,921 expert reproduction requirements in PaperBench found only 45.4% fully documented in the original PDFs. A separate failure-cost analysis across 24,008 AI agent runs found 90.2% of compute was spent on failed exploration that traditional papers throw away. The framing put to IEEE Spectrum is direct: 350 years after the scientific paper replaced secrecy with peer-reviewed archives, the PDF has become "lossy compression" of the research process, charging a "storytelling tax" on discarded branches and an "engineering tax" on incomplete specifications.
The proposers flag their own limits. ARA has only been validated in code-reproducible ML; wet-lab, theoretical, and field-science workflows are out of scope. The protocol has no fine-grained access control, sandboxing, or anomaly detection yet. In the reproduction run, ARA produced 1 hallucinated result out of 15 versus 2 out of 15 in the legacy group. The preprint also names a path-dependence risk: capable agents may converge on a single working solution and discard alternatives that would have surfaced in human exploration.
The tooling is shipping. The ARA-Labs GitHub repository bundles six agent skills (research-manager, compiler, rigor-reviewer, research-visualizer, research-foresight, and submit-ara) plus a one-line installer. NeurIPS 2026 has a workshop on AI-driven research ecosystems, and the team is building an ARA Hub for hosted artifacts. Adoption is still forward-looking: no major venue has yet required ARA-style submission, and independent replication of the benchmark gains is the open task.
A PDF asks the reader to take the prose as the source of truth and the code as a courtesy. An ARA asks the reader to take the executable artifact as the source and the prose as a guided tour. The 350-year-old scientific paper is not being retired this quarter, but the question put to IEEE Spectrum — "do AI scientists need a new research paper format?" — now sits next to a working code repo, measured deltas, and a list of what it cannot yet do.