This paper introduces OmniScientist, an AI system that tries to do end-to-end scientific work starting from raw evidence. Instead of only reading text or precomputed tables, the system perceives images, waveforms, video, 3‑D structures, tables, formulas and more. The authors say this lets the AI form and test questions that depend on spatial, temporal and cross-channel details that scalar summaries can hide.

OmniScientist combines a perception layer with three autonomous agents that handle idea generation, experiments, and writeup. The agents run inside a deterministic pipeline that moves a study from observation to claims. At each stage the system loops between observing, reasoning, and acting so that what it sees can change the question, the experiment design, and the final statements. The pipeline also runs automated idea, rigour, and claim checks in code to screen for novelty, enforce statistical validity, record how results were produced (execution provenance), and keep numerical steps traceable.

The authors tested the system on a 36‑case demonstration suite. These cases cover five broad discipline families and four families of scientific evidence, and use many modalities such as images, signals, audio, video, trajectories, tables, formulae and graphs. OmniScientist completed the full path from raw data to a compiled manuscript for all 36 cases. With the reference reasoning backbone the generated papers received a mean overall score of 6.3 on a seven‑point rubric scored by two cross‑field judges. When compared to a blind variant that only saw precomputed scalar features, direct perception improved every evaluated dimension and won 85% of head‑to‑head judgments.

Why this matters: many current AI research agents work only from text, labels, or numbers that humans prepare first. The paper argues that keeping raw observations available across the whole research lifecycle preserves spatial, temporal and procedural relations that can be decisive for discovery. The system is discipline‑agnostic in design: adding a new domain needs only a new specification file, not a rewrite of the core pipeline.