OmniScientist: An Omni-Modal Omni-Discipline AI Scientist
The paper says direct perception of raw scientific evidence improved every evaluation dimension over a scalar-feature-only variant.
OmniScientist is presented as an end-to-end AI scientist that works from heterogeneous raw evidence across images, signals, audio, video, 3-D structures, trajectories, tables, formulae, and graphs. Its pipeline uses a perception layer plus agents for ideation, experiments, and writeup, with code-based checks for novelty, statistics, provenance, and numerical traceability. The authors report tests on 36 real-data cases across five discipline families, where the system produced compiled manuscripts in every case. In paired comparisons, the perception-enabled system won 85% of head-to-head judgments. HF Daily Papers' note
OmniScientist is presented as an end-to-end AI scientist that works from heterogeneous raw evidence across images, signals, audio, video, 3-D structures, trajectories, tables, formulae, and graphs. Its pipeline uses a perception layer plus agents for ideation, experiments, and writeup, with code-based checks for novelty, statistics, provenance, and numerical traceability. The authors report tests on 36 real-data cases across five discipline families, where the system produced compiled manuscripts in every case. In paired comparisons, the perception-enabled system won 85% of head-to-head judgments. HF Daily Papers' note
score 5