OmniScientist Explained: How an AI Learns to Observe the World Like a Human Scientist
This is an English rendering of a Chinese forum post that explains the OmniScientist paper in a Feynman-style, narrative format.
The Problem: Blindfolded AI Scientists
Most current "AI scientist" systems—such as Sakana AI's AI Scientist, or LLMs paired with toolchains—can read papers, write code, and generate hypotheses, but they never see raw evidence. They work from preprocessed numbers, summaries, and labels produced by humans or other models. The author likens them to a brilliant scientist who can only receive notes slipped under the door, never directly observing microscope images, seismic signals, or simulation trajectories.
OmniScientist's central thesis: real scientific discovery often begins with direct perception of phenomena, not second-hand processing.
Giving AI the Five Senses
OmniScientist's Perception Layer directly ingests raw evidence across essentially all scientific data forms:
- Images (microscopy, astronomy, medical imaging)
- Signals and audio (EEG, seismic waves, spectra)
- Video of dynamic processes
- 3D structures (molecules, protein folds, point clouds)
- Trajectories, tables, formulas, and graph networks
- From tool to executor: AI may run the full research chain, with humans setting directions and judging findings—amplifying, not replacing, scientists.
- Cross-disciplinary discovery: one system can observe protein structures and nebulae alike, surfacing hidden analogies.
- Democratization: reliable automated analysis of raw multimodal data lowers the barrier for resource-limited researchers—though guarding against "AI-generated junk science" remains an open challenge.
- Li, B., Fei, H., Ju, T., Lee, M. L., & Hsu, W. (2026). OmniScientist: An Omni-Modal Omni-Discipline AI Scientist. arXiv preprint arXiv:2608.13558.
- Sakana AI. (2024). The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery. arXiv:2408.06292.
- Feynman, R. P. (1965). The Character of Physical Law. MIT Press.
- Krenn, M., et al. (2022). Scientific discovery in the age of artificial intelligence. Nature, 620(7972), 47-60.
- Wang, H., et al. (2024). Large Language Models for Scientific Discovery: A Survey. arXiv:2402.10091.
A concrete example: analyzing electron microscopy of a novel superconductor. A traditional pipeline has humans extract features first—an irreversible set of prior judgments that could filter out genuinely novel structures. OmniScientist instead examines the full raw pixel space and autonomously discovers patterns and anomalies.
Three Agents in a Deterministic Pipeline
OmniScientist automates the research lifecycle via three cooperating agents sharing one perception "brain":
1. Ideation Agent – formulates novel, feasible, important, and testable research questions; each idea passes an automated Idea Check (executed as code to verify basic logic). 2. Experiment Agent – designs experiments, runs statistics, and enforces a Rigour Check: multiple-comparison correction, pre-registered analysis plans, traceable intermediate steps, and execution provenance—explicitly guarding against p-value hacking. 3. Writeup Agent – builds a rigorous argument chain; Claim Check ensures every major claim is backed by data or derivation, including honest reporting of limitations.
Evaluation: 36 Cases, 5 Disciplines
The system was tested on 36 real-data cases across five disciplines (biomedical, physical, earth/environmental, engineering, social sciences) and four evidence types (spatial, temporal, relational, procedural). All cases completed the full data-to-paper pipeline with an average score of 6.3/10—roughly the quality of a solid paper needing minor revisions.
The key ablation: the version directly perceiving raw evidence beat the feature-only version on all 7 evaluation dimensions, winning 85% of head-to-head comparisons. Direct perception is a qualitative leap, not a nicety.
Why It Matters
Limitations
A 6.3/10 average shows plenty of headroom. Nobel-level breakthroughs still demand deep domain intuition, taste for important problems, boldness under uncertainty, and resilience—qualities current AI lacks. But OmniScientist demonstrates that the first step—AI directly perceiving the world—has been taken, echoing Feynman's "What I cannot create, I do not understand."