Is This the Last Human-Written Paper? Converting PDFs into AI-Executable Research Artifacts

For three centuries the scientific community has communicated results through static PDF papers, a format that compresses the messy, iterative nature of research into a polished narrative.

The new arXiv paper The Last Human-Written Paper: Agent-Native Research Artifacts (arXiv:2604.24658), authored by 37 researchers from Stanford, MIT, CMU, Michigan and others, asks whether this paradigm still makes sense when both authors and readers are AI agents.

The authors identify two "invisible taxes" imposed by PDFs:

Storytelling Tax : failed experiments, pivots and dead‑ends are omitted, causing irreversible information loss for agents that could otherwise reuse that knowledge.

Engineering Tax : essential details such as hyper‑parameters, warm‑up schedules and low‑level tricks are often missing, creating a gap between "sufficient to convince reviewers" and "sufficient to execute".

Using PaperBench’s 8,921 expert‑annotated reproducibility requirements, they quantify the problem: only 45.4% of needed information is fully described, 26.2% of hyper‑parameters are absent, 21.9% are vague, 13.4% rely solely on cross‑references, and 21.7% lack code or baseline details.

To eliminate these taxes, they propose the Agent‑Native Research Artifact (ARA), a structured package consisting of four interlocking layers:

Cognitive Layer : formal claims, concepts and declarative experiment designs.

Physical Layer : ready‑to‑run code and environment specifications.

Exploration Graph Layer : a directed‑acyclic graph preserving every pivot, dead‑end and alternative path.

Evidence Layer : direct links from each claim to raw experimental outputs.

Three mechanisms make ARA practical without extra researcher effort:

Live Research Manager : silently records decisions, heuristics and failures during AI‑human co‑research, automatically constructing the artifact.

ARA Compiler : translates legacy PDF + code repositories into ARA format, preserving existing literature.

ARA‑native Review System : automates objective checks (e.g., hyper‑parameter reporting, evidence linkage), allowing human reviewers to focus on novelty and significance.

Experimental evaluation on PaperBench and RE‑Bench compares ARA with the traditional "PDF + GitHub" bundle for AI agents. Results show:

Understanding : accuracy rises from 72.4% to 93.7% (+21.3 pp) across 450 benchmark questions.

Reproduction : success rate improves from 57.4% to 64.4% (+7.0 pp) on 15 papers and 150 sub‑tasks; gains are larger for harder tasks.

Extension : ARA wins 3 of 5 open‑ended RE‑Bench tasks and enables agents to take useful first actions earlier.

A notable tension emerges: for very strong agents, retaining all dead‑ends can constrain exploration, suggesting a need for a "forgetting" mechanism when agents exceed a certain capability.

Overall, the study concludes that when AI agents become the primary consumers of research, packaging papers as structured ARA artifacts yields far better comprehension, reproducibility and extensibility than the conventional PDF + code approach.