Poster
in
Workshop: AI as a Tool for Mathematics, Computer Science, and Machine Learning Thu, Jul 9, 2026 • 7:50 PM – 9:00 PM PDT

Agent-Native Research Artifacts: A Reproducible Workflow for AI-Assisted Research

Jiachen Liu ⋅ Jiaxin Pei ⋅ Jintao Huang ⋅ Chenglei Si ⋅ Ao Qu ⋅ Robert Tang ⋅ Runyu Lu ⋅ Lichang Chen ⋅ Xiaoyan Bai ⋅ Haizhong Zheng ⋅ Shirui Chen ⋅ Zhiyang Chen ⋅ Haojie Ye ⋅ Yujuan Fu ⋅ Zexue He ⋅ Zijian Jin ⋅ Zhenyu Zhang ⋅ Shangquan Sun ⋅ Maestro Harmon ⋅ Dianzhuo Wang ⋅ Qian-Ze Zhu ⋅ Jiachen Sun ⋅ Mingyuan Wu ⋅ Baoyu Zhou ⋅ Chenyu You ⋅ Shijian Lu ⋅ Yiming Qiu ⋅ Fan Lai ⋅ Yuan Yuan ⋅ Yao Li ⋅ Junyuan Hong ⋅ Ruihao Zhu ⋅ Beidi Chen ⋅ Alex Pentland ⋅ Ang Chen ⋅ Mosharaf Chowdhury ⋅ Zechen Zhang

Project Page [ Slides] [ Poster] [ OpenReview]

Abstract

AI-assisted research generates rich process knowledge as a byproduct of every researcher agent session: the decisions weighed, the dead ends explored, the hyperparameter rationale, the heuristics that finally converged. The conventional artifact (a PDF and a code repo) captures only a sliver of this and forces the next user, often another AI agent, to reverse-engineer the rest. We present a reproducible workflow that turns ordinary AI-assisted research into a structured, agent executable artifact, the Agent-Native Research Artifact (ARA), with three coordinated coding agent skills: the Live Research Manager (LRM) silently captures research-significant events at session boundaries, the ARA Compiler converts legacy PDFs and repos into the same format, and the ARA Seal pipeline verifies the result before downstream use. Each skill is a natural language specification loadable by any coding agent (Claude Code, Cursor, etc.) in seconds. On 30 ML papers (PaperBench and RE-Bench), an agent operating on the resulting artifact answers questions at 93.7% vs. 72.4% on the PDF baseline and reproduces methods at 64.4% vs. 57.4%. We discuss the cognitive task each skill solves, walk through this paper’s own ARA as a self applied worked example (the appendix shows the resulting /logic, /trace, and /evidence layers), and document the failure modes we have hit.

Chat is not available.