The Last Human-Written Paper: Agent-Native Research Artifacts

Abstract

Traditional research papers discard failed experiments and obscure implementation details, hindering reproducibility and cumulative scientific progress. We propose Agent-Native Research Artifacts (Ara), a machine-executable format that replaces the static paper with a structured, agent-consumable representation of a research project. Ara comprises four layers—scientific logic, executable code, exploration graphs, and evidence records—supported by a Live Research Manager, an Ara Compiler, and an automated review system. Across benchmark evaluations, Ara improves question-answering accuracy from 72.4% to 93.7% and substantially raises reproduction success rates, suggesting a viable path toward research artifacts designed for both human and agent readers.

Publication
arXiv preprint
Junyuan "Jason" Hong
Junyuan "Jason" Hong
Incoming Assistant Professor

My research interest lies in the interaction of responsible AI and healthcare.

Related