OmniScientist
OmniScientist (Li et al., 2026) is an end-to-end, omni-modal AI scientist that conducts multidisciplinary research directly from heterogeneous raw evidence — images, signals, audio, video, 3-D structures, trajectories, tables, formulae, and graphs — producing a finished scientific manuscript from raw data with no human in the loop. Developed at the National University of Singapore (Li, Fei, Ju, Lee, Hsu) with Oxford collaboration.
arXiv: 2608.13558 — https://arxiv.org/abs/2608.13558 Project: https://omni-scientist.github.io Code: https://github.com/Omni-Scientist/OmniScientist
Core Architecture
The system has two layers:
- Perception layer — directly reads raw multimodal evidence (photographs, spectrograms, waveforms, volumes, point clouds, video, trajectories, tables, sequences) across 20+ disciplines without first flattening every research object into a scalar benchmark.
- Three autonomous agents operating within a deterministic pipeline:
- Ideation agent — generates research questions shaped by observations
- Experiment agent — designs and runs experiments, makes decisions based on perceived evidence
- Writeup agent — produces the final manuscript
The pipeline runs: Perceive → Hypothesis → Experiment → Report → Feedback, with observations flowing through the entire lifecycle — not just at input.
Rigour Enforcement
Three categories of automated checks, all run in code:
- Idea checks — novelty screening against existing literature
- Rigour checks — statistical validity and execution provenance
- Claim checks — numerical traceability (every number in the paper must trace back to a code execution)
When a check fails, the system is sent back to an earlier stage. The interactive demos show 3–6 “sent back” events per run.
Evaluation
- 36 real-data cases spanning 5 discipline families and 4 evidence families — all completed the full path from raw data to compiled manuscript.
- Mean overall paper score: 6.3 (with the reference reasoning backbone).
- Blind comparison: Against a variant receiving only precomputed scalar features, direct perception improves all 7 evaluation dimensions and wins 85% of head-to-head judgments — demonstrating that lifecycle-wide perception is essential for evidence-grounded discovery.
Backbone Benchmark
A cross-family judge panel scored papers 0–10 across 7 dimensions. GLM-5.2 achieved the highest composite (6.6), followed by Sonnet 5 and Kimi K2.7 (both 6.5). Smaller open models (Gemma-4-9B: 4.1) lag substantially.
Five Representative Papers
| Paper | Discipline | Evidence | Score |
|---|---|---|---|
| Sequential vs bursty leaf initiation in 3-D plant scans | Plant science | 3-D scan | 7.2 |
| Cramér-Rao scaling of exponent precision in monomial Feynman laws | Physics | Formula | 7.1 |
| Transient-impulsivity features across machine types | Machinery | Audio | 7.1 |
| Coherent polarized signals in noise-labelled STEAD traces | Seismology | Signal | 6.9 |
| A continuum-removed index for residue and tillage ordering | Remote sensing | Hyperspectral | 6.3 |
Representative Findings
OmniScientist produced several non-trivial scientific discoveries:
- Seismology: 21.7% of noise-labelled STEAD traces carry coherent polarized transients — the finding depends on cross-channel polarization, not amplitude alone (which flags only 2.0%).
- Cardiology: A heart-sound classifier’s apparent performance (AUC 0.60) was largely a cohort shortcut — leave-one-cohort-out drops it to 0.35 (below chance), while the genuine acoustic murmur feature is stable (0.701 → 0.656).
- Astronomy: Galaxy morphology classifications from a general-purpose VLM are consistent across deep (SDSS) and shallow (DECaLS) surveys (83.8% agreement, κ = 0.75).
- Materials: Random-split cross-validation inflates superconductor Tc prediction by 3.1–7.0× compared to leave-one-family-out — the model interpolates within families rather than generalizing.
- Remote sensing: A land-cover benchmark’s performance depends on frame-locked orientation cues; rotating patches collapses accuracy from 62.0% to 50.7%.
Key Distinction from Other AI Scientists
Most AI scientist systems reason over text, code, labels, or precomputed summaries. OmniScientist’s central claim is that workflow coverage alone is insufficient — the agent must perceive raw evidence directly throughout the research lifecycle. The 85% win rate over the blind variant is the key evidence for this claim. This positions OmniScientist as the first system to demonstrate that omni-modal perception is essential, not merely beneficial, for evidence-grounded scientific discovery.
Related Pages
- ai-scientist — curated hub of autonomous AI research systems (The AI Scientist, Co-Scientist, DeepEvidence, etc.)
- deepevidence — biomedical deep research agent; complementary domain-specific approach to evidence synthesis
- what-is-intelligence-book — broader question of whether prediction-based AI constitutes genuine intelligence