OmniScientist: An Omni-Modal AI Scientist for Multidisciplinary Research

What changed

Researchers have developed OmniScientist, an AI system capable of conducting multidisciplinary research by directly processing heterogeneous raw evidence across multiple modalities. Unlike previous AI scientists that primarily reason over text, code, or precomputed summaries, OmniScientist incorporates a perception layer and three autonomous agents (ideation, experiment, write-up) that operate within a deterministic pipeline. This architecture allows raw observations from various data types to directly inform research questions, experimental decisions, and final claims throughout the research lifecycle. The system also includes built-in checks for novelty, statistical validity, execution provenance, and numerical traceability.

Why it matters for builders

This work advances the capabilities of AI scientists by enabling them to handle a wider range of scientific evidence, including images, signals, audio, video, 3-D structures, trajectories, tables, formulae, and graphs. For AI builders, this signifies a move towards more comprehensive and autonomous research agents that can operate with a deeper understanding of complex, real-world data, moving beyond text-based reasoning.

Practical impact

OmniScientist was evaluated on 36 real-data cases across five discipline families and four families of scientific evidence. The system successfully completed the full research path from raw data to a compiled manuscript in all cases. In direct comparisons, OmniScientist's direct perception approach improved all seven evaluation dimensions and won 85% of head-to-head judgments against a variant that only received precomputed scalar features. This demonstrates the value of lifecycle-wide perception for evidence-grounded scientific discovery.

Caveats and source limits

The provided source is a research paper detailing the OmniScientist system and its evaluation. Specific implementation details, code availability, and performance benchmarks beyond the reported evaluation metrics are not detailed. The reported results are based on the authors' evaluation within the scope of their research.