Goodbye PDF? Stanford et al. Propose ARA Format for AI Research Agents
量子位 · wechat · 2026-08-09
A team of 37 researchers from Stanford, UMich, CMU, and MIT has introduced ARA (Agent-Native Research Artifact), urging the academic community to abandon traditional PDF papers in favor of an AI-centric research knowledge package.
- Limitations of PDF: Traditional papers prune failed paths for narrative coherence ("narrative tax") and omit crucial environment and parameter details ("engineering tax"), making reproduction difficult for AI. Data shows only 45.4% of reproduction requirements in PaperBench are fully specified in PDFs.
- ARA Architecture: ARA restructures research into a machine-operable knowledge package containing scientific logic, executable code, an exploration graph (preserving failed paths), and an evidence layer.
- Performance: In comprehension tests, AI using ARA achieved 93.7% accuracy (vs. 72.4% for traditional materials). In extending research, ARA helps AI quickly bypass invalid directions by leveraging failure records, saving significant exploration costs.
More from coding & agent
- DIRECT: retrieval-based next-question suggestions for conversational agents — burkov · 2026-10-16
- Claude Code and Codex now autocomplete the user's next action — tokenbender · 2026-10-16
- Hermes Agent ships official Reverse Engineer Anything plugin with one-command setup — Teknium · 2026-10-16
- MLExpert: 35B MoE Agent Trained to Autonomously Handle Machine Learning Engineering — Shangheng Du · 2026-10-16
- Cowork-Gym: Alibaba Synthesizes 64.9K Agent Work Environments, Boosting Qwen by 8.7 Points — Anni Zou · 2026-10-16
- MMSearch-OPD: Joint RL and On-Policy Distillation for Long-Horizon Multimodal Search Agents — PekingUniversity · 2026-10-16