PDF Should Die, ARA Should Rise: Time for Agent-Native Research Papers

A team of 37 researchers from Stanford, University of Michigan, Carnegie Mellon, and MIT, led by first author Liu Jiachen, proposes replacing traditional PDF papers with ARA (Agent-Native Research Artifacts). They argue that PDFs impose a 'narrative tax' and 'engineering tax,' losing exploratory paths and critical details. ARA reorganizes research into four layers: scientific logic, executable code, exploration map, and evidence, making it directly operable by AI. In tests with 450 questions, AI using ARA achieved 93.7% accuracy vs 72.4% with traditional materials; reproduction success was 64.4% vs 57.4%. ARA also helped AI avoid failed directions early, saving costs, as 90.2% of AI agent run costs are wasted on failed exploration. However, ARA is not yet widely applicable, has privacy and security gaps, and experienced one hallucination instance. The team designed three mechanisms: Live Research Manager, ARA Compiler, and an ARA-native review system. Liu founded Agent-Native Research Lab in May 2026 to promote the ARA Commons ecosystem.