This might be the last paper humans ever write for AI to read. I recently came across a paper co-authored by 37 researchers from Stanford, CMU, Michigan, and others, titled The Last Human-Written Paper. The core argument is sharp: the paper format that has been used for centuries may be obsolete in the age of AI. The authors point out two “invisible taxes” that have long been overlooked.

One is the storytelling tax: to craft a pretty story, we delete failed experiments, dead ends, and overturned hypotheses. What AI reads is a “walkthrough guide,” but it misses the truly valuable “pitfall records.” The other is the engineering tax: implementation details in a paper are usually enough to convince reviewers, but insufficient for an agent to directly reproduce. Many key tricks remain hidden in the author’s mind, code comments, and Slack logs.

So the authors propose ARA, which directly transforms the paper into a “research package” that agents can read and execute: it doesn’t just tell you the conclusions, but packages how they got there, how the code runs, where the evidence chain is, and which paths didn’t work.

I think the most interesting part of this paper is that it doesn’t discuss how AI can help humans write papers, but asks: when AI becomes a reader and executor of papers, should papers still look the way they do today? The core output of future research may no longer be “how much it looks like a paper,” but whether it can be understood, reproduced, tracked, and further extended by AI.

Humans have been writing papers for centuries. Next, we may start writing research packages for agents to execute. https://arxiv.org/pdf/2604.24658

1 Introduction

Source: https://arxiv.org/html/2604.24658 The Last Human-Written Paper: Agent-Native Research Artifacts Jiachen Liu1, *, Jiaxin Pei2, Jintao Huang3, Chenglei Si2, Ao Qu4, Xiangru Tang5, Runyu Lu1, Lichang Chen6, Xiaoyan Bai7, Haizhong Zheng8, Carl Chen9, Zhiyang Chen10, Haojie Ye11, Yujuan Fu12, Zexue He2, Zijian Jin13, Zhenyu Zhang2, Shangquan Sun14, Maestro Harmon15, Dianzhuo Wang16, Qian-ze Zhu16, Jianqiao Zeng, Jiachen Sun17, Mingyuan Wu18, Baoyu Zhou19, Chenyu You20, Shijian Lu14, Yiming Qiu21, Fan Lai18, Yuan Yuan22, Yao Li23, Junyuan Hong24, Ruihao Zhu25, Beidi Chen8, Alex Pentland2, Ang Chen1, Mosharaf Chowdhury1, Zechen Zhang16, 15 1University of Michigan,2Stanford University,3Ohio State University,4MIT,5Yale University,6Meta Superintelligence Labs,7University of Chicago,8Carnegie Mellon University,9University of Washington,10University of Toronto,11NVIDIA,12Meta,13New York University,14Nanyang Technological University,15Orchestra Research,16Harvard University,17LinkedIn,18UIUC,19Arizona State University,20Stony Brook University,21University of Hong Kong,22Boston College,23Portland State University,24National University of Singapore,25Cornell University

Scientific publication compresses a branching, iterative research process into a linear narrative, discarding the majority of what was discovered along the way. This compilation imposes two structural costs: a Storytelling Tax, where failed experiments, rejected hypotheses, and the branching exploration process are discarded to fit a linear narrative; and an Engineering Tax, where the gap between reviewer-sufficient prose and agent-sufficient specification leaves critical implementation details unwritten. Tolerable for human readers, these costs become critical when AI agents must understand, reproduce, and extend published work. We introduce the Agent-Native Research Artifact (Ara), a protocol that replaces the narrative paper with an agent-executable research package structured around four layers: scientific logic, executable code with full specifications, an exploration graph that preserves the failures compilation discards, and evidence grounding every claim in raw outputs. Three mechanisms support the ecosystem: a Live Research Manager that captures decisions and dead ends during ordinary development; an Ara Compiler that translates legacy PDFs and repos into Ara s; and an Ara -native review system that automates objective checks (analogous to a grammar checker for prose) so human reviewers can focus on significance, novelty, and taste. On PaperBench and RE-Bench, Ara raises question-answering accuracy from 72.4% to 93.7% and reproduction success from 57.4% to 64.4%. On RE-Bench’s five open-ended extension tasks, preserved failure traces in Ara accelerate progress, but can also constrain a capable agent from stepping outside the prior-run box depending on the agent’s capabilities. Correspondence: Jiachen Liu ([email protected] (https://arxiv.org/html/2604.24658v3/mailto:[email protected])) Code: github.com/AmberLJC/Agent-Native-Research-Artifact (https://github.com/AmberLJC/Agent-Native-Research-Artifact) ARA Commons \icml@noticeprintedtrue

Research produces a rich, branching knowledge object: months of hypotheses tested and rejected, implementation tricks discovered through trial and error, design alternatives weighed against each other, and the full exploration trajectory that explains why the final approach was chosen. Publishing compiles this object into a linear narrative (Medawar,1963 (https://arxiv.org/html/2604.24658#bib.bib62); Canini,2026 (https://arxiv.org/html/2604.24658#bib.bib6)), discarding failed experiments, tacit engineering knowledge, and the branching process to satisfy the conventions of human-readable storytelling (Rosenthal,1979 (https://arxiv.org/html/2604.24658#bib.bib65); Francoet al.,2014 (https://arxiv.org/html/2604.24658#bib.bib66)). This compilation cost, a consequence of the documentation convention rather than any particular file format, was tolerable when every consumer of a paper was human. It is not when AI agents routinely read papers to understand a field, reproduce experiments to validate findings, and extend published methods to new settings (Luet al.,2024 (https://arxiv.org/html/2604.24658#bib.bib64); Liu,2026a (https://arxiv.org/html/2604.24658#bib.bib26)): each task requires precisely the knowledge that compilation discards (Figure1 (https://arxiv.org/html/2604.24658#S1.F1)). More specifically, the compilation incurs two structural costs.

Refer to caption Figure 1: Publishing compiles a rich research object into a lossy narrative (left); Ara preserves the original as a high-fidelity, agent-executable knowledge package (right).

The first is the Storytelling Tax: the systematic erasure of research process knowledge imposed by compilation into narrative (Figure2 (https://arxiv.org/html/2604.24658#S1.F2)). Research does not proceed linearly—it branches, backtracks, and accumulates hard-won failure knowledge before converging on a publishable result (Kuhn,1962 (https://arxiv.org/html/2604.24658#bib.bib61); Medawar,1963 (https://arxiv.org/html/2604.24658#bib.bib62)). Narrative compilation flattens this process into a polished linear story, discarding every failed experiment, rejected hypothesis, and abandoned approach. This emphasis on success leaves failures undocumented; although modern platforms archive final artifacts, the branching research process remains unrecorded, causing independent rediscovery of the same dead ends across groups (Rosenthal,1979 (https://arxiv.org/html/2604.24658#bib.bib65); Francoet al.,2014 (https://arxiv.org/html/2604.24658#bib.bib66)). Our own analysis of the METR eval-analysis-public dataset (Wijket al.,2025 (https://arxiv.org/html/2604.24658#bib.bib79)), covering 24,008 agent runs across 21 frontier models on RE-Bench, quantifies the cost (per-task breakdown in AppendixE.3 (https://arxiv.org/html/2604.24658#A5.SS3)): failed runs account for 90.2% of total dollar cost (and 59.2% of tokens), with a median failed-to-success token ratio of 113× \times, agents without access to prior failure records must independently rediscover every dead end. Equally lost is the record of human judgment along the trajectory: every rejection, revision, and endorsement is a preference signal over what constitutes good research, the scarce resource that binds once agents shoulder the grunt work. Narrative compilation discards this signal; a preserved trajectory renders it as structured supervision that compounds across projects.

The second is the Engineering Tax: the gap between reviewer-sufficient and agent-sufficient documentation (Figure3 (https://arxiv.org/html/2604.24658#S1.F3)). The paper communicates its contribution at the level of detail needed to convince a human reviewer; the codebase provides an implementation but not the operational specification needed to execute it. Between the two lies tacit knowledge (Polanyi,1966 (https://arxiv.org/html/2604.24658#bib.bib63))—algorithmic tricks, implementation decisions, and configuration choices—knowledge that exists in no written document and is transmitted only through direct lab contact or painstaking reverse-engineering. We quantify this void by classifying each of PaperBench’s 8,921 expert-annotated reproduction requirements across 23 ICML 2024 papers (Staraceet al.,2025 (https://arxiv.org/html/2604.24658#bib.bib18))against its source PDF (per-category breakdown and gap-type taxonomy in AppendixE.2 (https://arxiv.org/html/2604.24658#A5.SS2)): despite widespread artifact sharing, only 45.4% are fully specified. Code development is the most underspecified category (37.3% sufficient), and missing hyperparameters alone account for 26.2% of all gaps (full breakdown in AppendixE.2 (https://arxiv.org/html/2604.24658#A5.SS2)): a fundamental mismatch between the precision at which papers are written (sufficient to produce belief) and the precision at which agents must operate (sufficient to produce correct execution) (Stoddenet al.,2016 (https://arxiv.org/html/2604.24658#bib.bib20); Baker,2016 (https://arxiv.org/html/2604.24658#bib.bib7)).

Both taxes have persisted throughout the history of research because the human reader has always been the bandwidth-limited layer processing a vast, non-linear research trajectory. Capable AI agents now offer a more efficient, human-proximal proxy that processes the trajectory at machine bandwidth on the reader’s behalf, and three trends suggest what such an artifact should look like.

First, AI agents have become indispensable research companions, co-authoring code, running experiments, and iterating on hypotheses alongside humans (Luet al.,2024 (https://arxiv.org/html/2604.24658#bib.bib64)), with LLM adoption associated with paper-production increases of 23.7%–89.3% across scientific fields (Kusumegiet al.,2025 (https://arxiv.org/html/2604.24658#bib.bib101)); the full research trajectory (every failure, implementation trick, configuration choice, design pivot) is now captured as machine-readable text in researcher-agent sessions, yet no protocol preserves it as a first-class output.

Second, humans and AI agents have divergent information needs: humans skim abstracts and figures (Renear and Palmer,2009 (https://arxiv.org/html/2604.24658#bib.bib102)), while agents benefit from exhaustive detail that strictly improves reproduction, verification, and extension, so a single artifact optimized for human narrative can no longer serve both.

Third, research is scaling into a massively parallel enterprise where agents fork, extend, and merge each other’s work at machine speed, shifting the bottleneck from individual productivity to artifact operability : narrative PDFs, compiled for sequential human reading, cannot be forked, diffed, or merged, but a structured, lossless artifact can, letting research compound like software.

Refer to caption Figure 2: The Storytelling Tax: research proceeds as a branching tree with dead ends (left), but publication compiles it into a linear narrative (right), discarding all failure knowledge.

Existing efforts address fragments of this problem. The FAIR principles (Wilkinsonet al.,2016 (https://arxiv.org/html/2604.24658#bib.bib19))mandate findable, accessible data but say nothing about the structure of research arguments . RO-Crate (Soiland-Reyeset al.,2022 (https://arxiv.org/html/2604.24658#bib.bib12))packages research artifacts as archival bundles, not executable objects. Nanopublications (Grothet al.,2010 (https://arxiv.org/html/2604.24658#bib.bib16))formalize atomic claims but lack the execution layer needed for reproduction. The emerging AGENTS.md standard (OpenAI,2025 (https://arxiv.org/html/2604.24658#bib.bib23))provides agent-oriented documentation for code repositories but does not address the epistemic structure of research itself. None of these efforts jointly structure scientific logic, executable code, and exploration history into a single operable object (§8 (https://arxiv.org/html/2604.24658#S8)).

We propose the Agent-Native Research Artifact (Ara), a protocol that recasts the primary research object from narrative document to agent-executable knowledge package, with papers serving as compiled views of the underlying artifact (Figure1 (https://arxiv.org/html/2604.24658#S1.F1)). Ara organizes research into four interlocking layers: structured scientific logic that distills the paper’s conceptual abstractions into queryable claims and dependency graphs; executable code with full operational specifications; an exploration graph preserving the branching research process—failed experiments, rejected hypotheses, and design pivots—that narrative compilation discards; and grounded evidence binding every claim to its raw empirical outputs. Instead of parsing prose, reverse-engineering repositories, and rediscovering dead ends, an agent operating on an Ara artifact can query structured claims, execute declarative specifications, and build on the full decision history directly—a research object designed not to be read, but to be operated .

To build the ecosystem around Ara , we develop three enabling mechanisms. The Live Research Manager (§3 (https://arxiv.org/html/2604.24658#S3)) captures research decisions and dead ends as natural side-effects of everyday development, producing conforming artifacts without additional documentation burden. The Ara Compiler (§4 (https://arxiv.org/html/2604.24658#S4)) translates legacy PDFs, repositories, and supplementary materials into Ara format, providing backward compatibility with the existing publication ecosystem. The Ara -Native Review System (§5 (https://arxiv.org/html/2604.24658#S5)) automates structural verification and budget-aware reproduction (analogous to a grammar checker for prose), redirecting expert attention from mechanical checking to judgment—significance, novelty, and taste (Aczelet al.,2021 (https://arxiv.org/html/2604.24658#bib.bib67)).

Refer to caption Figure 3: The reproduction information gap across 8,921 PaperBench requirements. (a)PDFs systematically under-specify code development tasks. (b)The three largest gap types are precisely the categories Ara ’s structured layers address.

Contributions. - •We identify two structural costs of compiling research into narrative—the

Storytelling Tax and the Engineering Tax —and introduce the Agent-Native Research Artifact (Ara): a protocol that recasts the primary research object from narrative document to agent-executable knowledge package organized into four interlocking layers (§2 (https://arxiv.org/html/2604.24658#S2)). - •We develop three enabling mechanisms: a Live Research Manager (§3 (https://arxiv.org/html/2604.24658#S3)) that captures research decisions during ordinary development; an Ara Compiler (§4 (https://arxiv.org/html/2604.24658#S4)) that translates legacy PDFs and repositories into Ara format; and an Ara -Native Review System (§5 (https://arxiv.org/html/2604.24658#S5)) that automates objective verification (analogous to a grammar checker for prose) so human reviewers can focus on judgment. - •We evaluate Ara across three layers of research utility (§7 (https://arxiv.org/html/2604.24658#S7)): understanding (what an agent can extract from the artifact), reproduction (whether the agent can re-execute the paper’s experiments), and extension (whether the agent can build beyond the documented results and discover genuinely new findings, the defining goal of research). Across all three layers, agents operating on an Ara consistently outperform those reading the paper PDF and its associated code repository.

2The

Ara Protocol

The Agent-Native Research Artifact (Ara) protocol defines a file-system ontology that transforms CS research from a narrative document into a machine-executable knowledge package. We describe the design philosophy (§2.1 (https://arxiv.org/html/2604.24658#S2.SS1)) and the layered architecture (§2.2 (https://arxiv.org/html/2604.24658#S2.SS2)). Figure6 (https://arxiv.org/html/2604.24658#S3.F6)(§3 (https://arxiv.org/html/2604.24658#S3)) illustrates how the Live Research Manager mediates between the human–AI research pro