AI 科研代理
OmniScientist Carries Multimodal Evidence Through the Research Workflow, With Direct Perception Winning 85% of Paired Evaluations
OmniScientist lets agents directly inspect images, signals, audio, video, 3D structures, and trajectories before completing ideation, experimentation, and paper writing. All 36 cases produced complete manuscripts, but the case count, code availability, and automated evaluation design still require clarification.
Existing “AI scientist” systems can often search the literature, write experimental code, and generate papers, yet they commonly interact only with textual descriptions, labels, or precomputed statistics. OmniScientist integrates raw images, waveforms, spectra, audio, video, 3D structures, trajectories, tables, equations, and graph data throughout the research lifecycle, preventing human-selected features from discarding spatial, temporal, or cross-channel relationships.
The system consists of a perception layer and three autonomous agents for ideation, experimentation, and writing, surrounded by a deterministic workflow. Each stage must pass programmatically executed checks covering novelty and falsifiability, data leakage, effective sample size, multiple comparisons, execution provenance, numerical traceability, and safeguards against revising hypotheses after inspecting results. If an experiment fails or produces a null result, the workflow can return to the ideation stage. New disciplines are added primarily through specification files that describe data and outputs, without requiring the core engine to be rewritten.
The paper reports 36 real-world data cases spanning five disciplinary categories and four evidence modalities, all of which progressed from raw data to compilable manuscripts. Using the reference reasoning model, the mean aggregate score across seven evaluation dimensions was 6.3/10. In paired comparisons against a blinded version that could access only budgeted scalar features, the direct-perception version improved across all seven dimensions and won 85% of judgments. The public project page also provides replays of three runs, execution logs, and generated manuscripts, allowing readers to inspect how the agents formed claims from evidence.
However, these scores were still produced primarily by model-based evaluators, and completing a paper is not equivalent to making a new peer-reviewed discovery. The paper reports 36 cases, while the project homepage lists 34. The homepage also claims that the code has been released, but at the time of writing, its GitHub link returned a 404 error. These versioning and reproducibility issues should be resolved before comparing costs, error-rollback capabilities, and genuine cross-laboratory generalization.