OmniScientist: An Omni-Modal Omni-Discipline AI Scientist
OmniScientist claims direct perception of raw scientific evidence improves an AI scientist’s results over feature-only inputs.
The arXiv paper describes a system with a perception layer and three agents for ideation, experiment, and writeup, arranged in a deterministic research pipeline. It was evaluated on 36 real-data cases across five discipline families and multiple evidence types, including images, signals, audio, video, 3-D structures, trajectories, tables, formulae, and graphs. The authors report that it completed a compiled manuscript in all 36 cases, with a mean overall paper score of 6.3 using the reference reasoning backbone. Against a blind variant given only precomputed scalar features, the direct-perception version improved all seven evaluation dimensions and won 85% of paired judgments. ArXiv · AI/CL/LG's note
The arXiv paper describes a system with a perception layer and three agents for ideation, experiment, and writeup, arranged in a deterministic research pipeline. It was evaluated on 36 real-data cases across five discipline families and multiple evidence types, including images, signals, audio, video, 3-D structures, trajectories, tables, formulae, and graphs. The authors report that it completed a compiled manuscript in all 36 cases, with a mean overall paper score of 6.3 using the reference reasoning backbone. Against a blind variant given only precomputed scalar features, the direct-perception version improved all seven evaluation dimensions and won 85% of paired judgments. ArXiv · AI/CL/LG's note
score 5