Mobus: Data Infrastructure for Researchers and Autonomous Scientific Agents

Harry Ilanyan, Hrant Virabyan

Track: Track 3: AI Scientist Proposal Competition

Keywords: AI Scientist, autonomous scientific agents, data infrastructure, dataset discovery, Model Context Protocol, license compliance

TL;DR: Mobus collapses days of dataset hunting into a single agent call and is building the universal data-infrastructure layer for science; discovery, manipulation and synthesis.

Abstract: Autonomous research agents increasingly retrieve and synthesize scientific data, but every step of the existing deep-research stack silently degrades the substrate of science: provenance is lost, license status is unverified, schema integrity is unchecked, and downstream work is unreproducible. We present Mobus, verifiable data infrastructure that treats provenance, license, and schema as first-class invariants preserved across iterative discovery, bounded refinement, and lineage-aware manipulation, producing analysis-ready datasets where every cell carries traceable, license-verified lineage. On a 104-question benchmark spanning eight scientific domains (biology, materials science, climate science, physics, public health, pure mathematics, advanced mathematics, computer science) and four baselines, Mobus achieves 95\% task completion versus 62\% for the strongest agentic baseline (GPT-Researcher), a 4\% license false-clean rate against 41\% to 58\% for retrieval-equipped baselines, 0.99 citation integrity, and 100\% correctness across 63 manipulation-tool unit tests, with judge-human agreement at Cohen's $\kappa = 0.715$.

Submission Number: 321