OmniScientist Runs Ideation, Experimentation and Manuscript Writing Across 36 Real Cases and Wins 85% of Head-to-Head Comparisons Against a Features-Only Baseline

arXiv (via HuggingFace Daily Papers)·low signalXBlueskyLinkedInCopy link

OmniScientist (arXiv 2608.13558, submitted August 13, by Bobo Li, Hao Fei, Tianjie Ju, Mong-Li Lee and Wynne Hsu) pairs a perception layer with three autonomous agents for research ideation, experimentation and manuscript composition, spanning images, signals, audio, video, 3-D structures, trajectories, tables, formulae and graphs. Across 36 real-world cases in 5 discipline families and 4 evidence families it completed the full path from raw data to compiled manuscript in every case, at a mean overall paper score of 6.3. The ablation is the useful result: direct perception of raw evidence beat a baseline restricted to precomputed features on all 7 evaluation dimensions and won 85% of head-to-head judgments, which argues against the common agent pattern of pre-summarizing inputs before the model sees them.