OmniScientist: An Omni-Modal Omni-Discipline AI Scientist
Bobo Li, Hao Fei, Tianjie Ju, Mong-Li Lee, Wynne Hsu
cs.AI, cs.CL
2026-08-14
OmniScientist, an omni-modal AI scientist from NUS and Oxford, perceives raw scientific evidence directly across the research lifecycle, completing all 36 tested cases and beating a scalar-only blind baseline in 85% of head-to-head comparisons.
Recent AI-scientist systems already automate an impressive stretch of the research pipeline, from generating hypotheses to running code to drafting a manuscript. But covering the workflow doesn't mean the system actually sees the evidence driving the science. Most existing systems reason over text, code, precomputed labels, or summary statistics, and the spatial, temporal, cross-channel, and procedural relationships that often determine the scientific conclusion get thrown away during that preprocessing: a caption can miss local morphology, an unordered feature vector can erase temporal order, a handful of scalars can hide a cross-channel inconsistency. OmniScientist, from a team at the National University of Singapore and Oxford, targets exactly this perception gap: building an AI scientist that works from heterogeneous raw evidence instead of someone else's pre-computed summary of it.
OmniScientist consists of a perception layer plus three autonomous agents (ideation, experiment, writeup) running inside a deterministic pipeline. The perception layer first sorts raw artifacts into four evidence families by the reasoning paradigm they require (perceptual, symbolic, quantitative-statistical, procedural), then into specific modalities, covering images, signals, audio, video, 3D structures, trajectories, tables, formulae, and graphs across 16 disciplines and 11 modalities in the test suite. The system prioritizes direct numeric analysis, extracting things like FFT peaks or trend points from raw records, and only renders a visualization when spatial or structural pattern matters, with a budget cap on visual inspection to avoid unbounded compute spend.
The ideation stage runs a ReAct loop: it inventories the available materials, decides whether to inspect raw observations, searches the literature through OpenAlex (with Crossref as a fallback), generates at least five candidate ideas, assesses novelty risk and feasibility for each, and finalizes the strongest one. That output has to pass a code-enforced check confirming it has a clear research question, hypothesis, experiment sketch, and falsification criterion, that it actually ran at least three focused literature searches and filtered five candidates, and that the study is executable in code rather than requiring a physical experiment. The system also automatically tempers overconfident claims, rewriting phrases like "first" or "never explored" into "appears under-explored based on this search."
The experiment stage translates the finalized idea into a methodological design and iterates through code generation, execution, and inspection inside a controlled runpython environment. The writeup stage selects and reports only claims traceable to the verified execution record, and formats the manuscript to match the target venue's conventions (machine-learning papers get Related Work and Limitations sections, biomedical papers put Methods last, chemistry papers merge Results and Discussion). All three stages share the same code-enforced idea, rigor, and claim checks, which require novelty screening, statistical validity, execution provenance, and numerical traceability, so every reported number has to map back to an actual program output.
The system completes the full pipeline, from raw data to a compiled manuscript, on all 36 cross-disciplinary cases (5 discipline families, 4 evidence families, and modalities spanning images, signals, audio, video, 3D, and trajectories). With Claude Sonnet 5 as the reasoning backbone, it achieves a mean overall score of 6.3. Switching backbones (GLM-5.2, GPT-5.6, Kimi K2.7, and several open-weight models) lands in a similar score range but with very different completion rates: Sonnet 5 finishes all 36 cases (mean composite 6.5), GLM-5.2 finishes 17 of the 18 cases it was dispatched (mean 6.7), GPT-5.6 finishes only 9 of 10, and the open-weight Qwen3.5-9B finishes just 18 of 32 (mean 4.1). Grouped by discipline or by modality, median composite scores stay tightly between 6.1 and 7.1, and scores correlate only weakly with manuscript length (Spearman ρ = 0.16), which rules out a simple write-more-words-score-higher shortcut.
Against a blind baseline that receives only precomputed scalar features, the perception-enabled system wins 85% of paired head-to-head comparisons across the five cases with a blind counterpart. The biggest gains are in multimodal grounding (+2.8) and significance (+1.8), with factual accuracy roughly tied between the two, meaning perception isn't buying more room to fabricate, it's buying access to claims a scalar-only system simply can't support. A leave-one-out ablation shows prior-art search matters most (removing it drops the mean score from 6.9 to 5.7, because the system starts proposing ideas that already exist in the literature); collapsing the iterative loop to a single pass also hurts quality, and removing the novelty check costs a full point on the novelty score.
Two case studies make the perception advantage concrete. Auditing roughly 1,500 three-component seismograms from the STEAD catalogue, the system spotted a clear onset-and-decay energy envelope inside a trace labeled noise, and instead of overriding the label, it turned the discrepancy into a testable question: what fraction of noise-labeled traces actually carry coherent transient energy. It built its own composite detector (an STA/LTA characteristic function plus amplitude, rectilinearity, planarity, and a cross-channel coincidence term) and found that 163 of 750 noise-labeled traces (21.7%, 95% CI [18.8%, 24.9%]) carry coherent, polarized cross-component bursts, an estimate that held up across a 50-fold range of false-alarm rates, three null models, four windowing choices, and a bootstrap over 417 stations. In a second case on pediatric chest radiographs, the system noticed pneumonic lungs weren't uniformly brighter but unevenly mottled, converted that observation into a spatial-dispersion metric on local Shannon entropy, and showed it separates normal from pneumonic cases with a large effect size (Cohen's d > 1.2) and a held-out AUC of 0.851 in a multivariable model. In both cases, the questions the perceiving system asked were ones a blind, scalar-only system's input interface simply couldn't surface.
The evidence here is that giving an AI scientist direct access to raw evidence, rather than someone else's summary of it, doesn't just make the writing better, it changes which questions the system can even ask and which claims it can support. In both case studies, the actual research question was invisible to a system limited to precomputed scalars. For anyone evaluating or building AI-scientist tooling, that's an easy-to-miss selection criterion: workflow coverage alone isn't the right bar, whether the system can actually touch raw data and whether it has code-level provenance checks against fabricated numbers matter just as much. The three-tier idea, rigor, and claim check design, paired with execution provenance, is also a concrete, reusable engineering pattern for keeping an LLM agent honest.
The paper doesn't include a standalone limitations section, but a few things stand out from the data itself. A mean score of 6.3 on a 0-10 scale is respectable but not high, and the paper doesn't compare these auto-generated manuscripts against real peer review outcomes. The quantified perception advantage is measured on only five paired cases (plus one vision-off control), a small sample for generalizing to all 36 cases or to disciplines and modalities beyond the tested suite. Scoring relies on two judge models from outside the tested systems (deepseek-v4-flash, gemini-2.5-flash-lite), so whatever biases or blind spots those judges have propagate into the final scores. Weaker open-weight backbones (Qwen3.5-9B, Gemma-4-26B) show noticeably lower completion rates, meaning the framework leans heavily on the capability of its reasoning backbone rather than working equally well with any small model dropped in.