OmniScientist: An Omni-Modal Omni-Discipline AI Scientist

Install:

◆Skill ▢Desktop ▶CLI

Paper:

OmniScientist — workbench

Not a mock-up. Every card is lifted from one recorded run — the request that was sent, the images it opened, the papers it read, the check that sent it back, the code it ran and the manuscript it produced.

Watch it work

Three complete runs, replayed from their recorded traces: the raw evidence it opened, the papers it read, every line of code it ran, and every time a check sent it back. Pause anywhere, drag the scrubber, take the paper and the log with you.

▶Open all three demos

Technical report · 2026

OmniScientist: An Omni-Modal Omni-Discipline AI Scientist

Bobo Li^♣ Hao Fei^♦^* Tianjie Ju^♣ Mong-Li Lee^♣ Wynne Hsu^♣

♣ National University of Singapore ♦ University of Oxford

Contact libobo@nus.edu.sg · haofei7419@gmail.com (* corresponding author)

10evidence modalities

20+disciplines

34end-to-end cases

News

2026/08Interactive replays of three complete runs are live: step through the evidence, the code and the checks, and take the paper and the log with you.
2026/07Technical report completed, covering the full OmniScientist evaluation.
2026/07Source code and the one-command research workflow are available on GitHub.
2026/07Five representative papers produced end to end by OmniScientist are listed with their scores.
2026/07Project page updated with results and multimodal evidence.

Raw evidence in, a finished study out

Observations of every kind go in; the engine reads them, tests what it reads, and writes the result up. Nothing on the way through is staged: the code runs against the real data, and a claim the numbers do not carry gets sent back rather than polished.

From raw observations through the engine to scientific discoveries

Which relations survive the interface

A caption or a feature vector arrives with the structure already removed. Three cases start at the raw record instead, and the cues read off it are what the hypothesis is built on. The band at the bottom is the interface most systems expose in its place.

Three cases from raw record to structural cues, hypothesis and verified finding, above the blind precomputed interface

One command, the whole research lifecycle

Evidence enters in its native form, and a deterministic pipeline routes the run through ideation, experiment and writeup, reporting only what the execution record supports. Perception stays available to every stage, and a new discipline takes a specification file, not a new engine.

The OmniScientist framework: raw evidence entering three agents for ideation, experiment and writeup, over a lifecycle-wide perception band

Only what the record carries becomes a claim

The rigour check asks whether the code really ran, whether every attempted test is accounted for, and whether the headline rests on an analysis that supports it. The claim check then matches every number and every sentence against the analysis that produced it.

Rigour check and claim check auditing a reported result against the execution record

Findings · two of these open a full write-up

Seismology · three-component signal

A fifth of the discard pile was never noise

21.7%163 of 750 noise-labelled traces carry coherent transients

See the full finding →

The detector never sees the label. Threshold calibrated on IAAFT surrogates of the noise itself at a 1% false-alarm rate, measured back at 1.07%.
An amplitude-only detector at the same rate flags 2.0%: polarization and cross-channel coincidence carry the finding.
Stable at 19 to 25% across a fifty-fold threshold range and three null models.

Eight traces labelled noise, four flat and four with clear impulsive onsets

Eight traces filed as noise. Lower half: the four with the strongest impulsive onsets in this sample.

Cardiology · audio

The metadata predicted the patient until the hospital changed

0.60 → 0.35metadata-only AUC, pooled then leave-one-cohort-out

See the full finding →

Below chance once the cohort is held out: the shortcut flips sign across sites. Drop 0.248, p < 0.0001.
The murmur-band acoustic feature barely moves: 0.701 to 0.656, not significant.
What sounded like a model of the heart was partly a model of the recording booth.

Heart sound spectrograms grouped by cohort

Heart-sound recordings, labelled normal or abnormal, grouped by the cohort that recorded them.

Astronomy · imaging

Change the telescope and its reading of the galaxy stays put

83.8%agreement on the same galaxies across two surveys

Chance agreement for this label mix is 35.0%; Cohen's κ is 0.75, CI 0.64 to 0.85.
Permutation test over the pairing gives p = 0.0002.
105 galaxies, each imaged twice: once by SDSS, once by DECaLS.

The same galaxies imaged by SDSS and DECaLS

Top row SDSS, bottom row DECaLS. Labels are the catalogue morphology.

Materials informatics · table

The usual split flattered the model by a factor of seven

3.1 - 7.0×error gap, random split versus leave-one-family-out

Cuprates: RMSE 12.1 → 51.6 K. Iron-based: 6.6 → 20.5 K. Other families: 4.3 → 30.3 K.
Same data, same model. Only the question changed, from interpolating inside a family to predicting one it had never seen.

Random split versus leave-one-family-out error

Figure produced by the run itself.

Remote sensing · imaging

Rotate the patches and the benchmark falls apart

62.0% → 50.7%orientation-only classifier, upright then rotated

Features carry no colour and no raw pixels, only orientation histograms, angular FFT and coherence. Chance is 20%.
Per class the collapse is uneven: forest recall 0.68 → 0.27, residential 0.91 → 0.62, annual crop does not move.
A frame-locked cue was doing part of the work that looked like land-cover understanding.

Sentinel-2 land-cover patches, six of the benchmark classes.

One engine, ten kinds of evidence

The same pipeline reads photographs, spectrograms, waveforms, volumes, point clouds, video, trajectories, tables and sequences, across more than twenty disciplines, without first flattening every research object into a scalar benchmark.

Reviewed by a cross-family judge panel

Scores run 0 to 10 across seven review dimensions. Composite is their mean. Blue marks the best in a column, green the second.

Backbone	Novelty	Sound.	Clarity	Signif.	Reprod.	MM ground.	Factual	Composite
GLM-5.2	6.2	7.1	6.8	6.4	5.9	6.6	7.5	6.6
Sonnet 5	6.3	7.0	7.0	6.3	6.1	5.1	7.8	6.5
Kimi K2.7	6.2	7.2	6.7	6.2	5.5	5.8	8.0	6.5
GPT-5.6	5.2	6.3	6.3	5.0	5.2	4.2	7.7	5.7
Qwen3.5-122B	4.7	5.5	6.2	4.8	4.8	4.8	6.5	5.3
Qwen3.5-27B	4.9	5.5	5.9	4.9	4.6	4.9	6.3	5.3
Gemma-4-31B	4.7	5.1	5.6	4.5	4.4	4.7	6.5	5.1
Gemma-4-26B	4.4	4.4	5.0	4.0	3.7	3.8	5.1	4.3
Qwen3.5-9B	4.0	4.1	4.8	3.7	3.7	3.9	4.8	4.1

Papers it wrote, start to finish

Paper	Discipline	Evidence	Score
Cramér-Rao scaling of exponent precision in monomial Feynman laws	Physics	Formula	7.1
Sequential versus bursty leaf initiation in 3-D plant scans	Plant science	3-D scan	7.2
A continuum-removed index for residue and tillage ordering	Remote sensing	Hyperspectral	6.3
Transient-impulsivity features across machine types	Machinery	Audio	7.1
Coherent polarized signals in noise-labelled STEAD traces	Seismology	Signal	6.9

@techreport{omniscientist2026,
  title  = {OmniScientist: An Omni-Modal Omni-Discipline AI Scientist},
  author = {Li, Bobo and Fei, Hao and Ju, Tianjie and Lee, Mong-Li and Hsu, Wynne},
  year   = {2026},
  note   = {Technical Report}
}

Every number on this page comes from the run that produced it. Observations are public research datasets; sources and licences are listed in the release documentation.

Watch it work

OmniScientist: An Omni-Modal Omni-Discipline AI Scientist

🎉News

Raw evidence in, a finished study out

Which relations survive the interface

One command, the whole research lifecycle

Only what the record carries becomes a claim

A fifth of the discard pile was never noise

The metadata predicted the patient until the hospital changed

Change the telescope and its reading of the galaxy stays put

The usual split flattered the model by a factor of seven

Rotate the patches and the benchmark falls apart

One engine, ten kinds of evidence

Reviewed by a cross-family judge panel

Papers it wrote, start to finish

News