AI
OmniScientist: an omni-modal three-agent scientific workflow in preprint

In a preprint submitted on 13 August and listed on arXiv on 14 August 2026, Bobo Li, Hao Fei, Tianjie Ju, Mong-Li Lee and Wynne Hsu describe a system in which ideation, experiment and write-up agents run in a deterministic pipeline with a perception layer.
The authors report that the manuscript path completed on 36 real-data cases, and that the system beat a scalar-feature-only variant in 85 per cent of head-to-head judgements.
The work is an architecture signal for evidence-aware research agents. But completing a manuscript is not scientific truth; independent replication, expert review and provenance remain essential.




