AI 代理／科學研究
OmniScientist Reads Raw Scientific Data Directly, Producing Traceable Papers Across 36 Cross-Disciplinary Tasks
OmniScientist carries perception through hypothesis generation, experimentation, and writing instead of first compressing images, waveforms, or trajectories into text summaries. In blinded comparisons, the full system achieved an 85% win rate against a version that received only precomputed features, though evaluations still relied primarily on model judges.
The newly released [OmniScientist](https://arxiv.org/abs/2608.13558) attempts to close an interface gap in automated scientific research agents. Although existing systems can search the literature, write code, and generate papers, they typically see only human-curated text, labels, or statistical features. Local image morphology, the temporal ordering of signals, and relationships across channels may therefore disappear before the research even begins.
The system divides evidence into four categories: perceptual, symbolic, quantitative-statistical, and procedural-dynamic. A perception layer supplies three agents—for ideation, experimentation, and writing—with raw data including images, audio, video, waveforms, 3D structures, trajectories, tables, formulas, and knowledge graphs. Each stage still uses a ReAct-style tool loop, but transitions through the workflow are controlled by deterministic programs: ideas must pass novelty and falsifiability checks; experiments must retain code, configurations, standard output, and figures; and the writing agent may cite only numbers traceable to execution records. If results are empty or an experiment fails, the workflow returns to the ideation stage instead of packaging the outcome as a discovery.
The authors ran the complete workflow on 36 real-world data cases spanning five disciplinary categories, including seismology, pathology, astronomy, agriculture, and engineering. Every case produced a compilable paper. The reference reasoning model's papers received an average score of 6.3. The direct-perception version outperformed a blinded version given only precomputed scalar features on all seven metrics, achieving an 85% win rate in pairwise evaluations. The [MIT-licensed repository](https://github.com/Omni-Scientist/OmniScientist) also provides a CLI, desktop workbench, skill packages, and five sample papers.
From an engineering perspective, the notable contribution is not simply another multi-agent pipeline, but the implementation of execution provenance, statistical constraints, and claim-to-evidence mappings as programmatic gates. However, the 36 cases constitute a demonstration suite rather than a preregistered competition. Paper quality was evaluated mainly by two cross-disciplinary model judges, and the work does not show that independent research teams can reproduce the novel findings produced by the system. The next steps should include blinded review by human experts, cross-laboratory reruns, and tests of whether the agents can maintain their constraints when raw data contain contamination or incorrect metadata.