Researchers have introduced OmniScientist, an artificial intelligence system designed to conduct multidisciplinary research using raw, heterogeneous evidence. Unlike previous AI systems that process text, code, or summaries, OmniScientist can directly interpret diverse data types, including images, signals, audio, video, 3-D structures, trajectories, tables, formulae, and graphs. The system aims to advance automated scientific discovery by processing a wider array of evidence.
OmniScientist operates through a deterministic pipeline involving a perception layer and three autonomous agents: one for ideation, another for experimentation, and a third for manuscript preparation. This structure allows observations from raw data to inform research questions, experimental decisions, and final claims throughout the entire research lifecycle. The system also incorporates checks for novelty screening, statistical validity, execution provenance, and numerical traceability, all enforced through code.
The researchers evaluated OmniScientist on 36 real-world data cases across five discipline families and four families of scientific evidence. In all cases, the AI system successfully completed the full research path, from raw data to a compiled manuscript. It achieved a mean overall paper score of 6.3 when using a reference reasoning backbone. In direct comparisons, OmniScientist outperformed a variant that relied solely on precomputed scalar features, improving all seven evaluation dimensions and winning 85% of head-to-head judgments. This suggests that the system's ability to perceive and process raw, multimodal data is critical for evidence-grounded scientific discovery.
Recent advancements in foundation models have enabled AI systems to automate more research workflows, including hypothesis generation, code execution, and manuscript writing. However, these systems often lack access to the full spectrum of evidence required for scientific discovery, as they typically only reason over text, code, labels, or precomputed summaries. This limitation leaves out scientifically decisive spatial, temporal, cross-channel, and procedural relations. OmniScientist addresses this gap by integrating a lifecycle-wide perception capability, which is presented as essential for broadly capable AI scientists.
The development of AI systems capable of assisting or automating scientific research is an active area of development. For instance, Stanford University researchers developed Biomni, an AI system that can read scientific papers, analyze datasets, write and execute code, form hypotheses, and design experiments. Biomni utilizes a wide array of tools and databases across various biological fields to accelerate research tasks. Google Research is also involved in developing AI co-scientists that can act as intelligent partners, offering guidance, generating ideas, and automating routine steps. These systems aim to augment human researchers, allowing them to focus on higher-level cognitive tasks.
Another project, also named OmniScientist, introduced a framework that simulates the human scientific system, including structured knowledge systems, collaborative research protocols, and open evaluation platforms. This earlier iteration aimed to foster a co-evolving ecosystem of human and AI scientists by enabling multi-agent collaboration and human researcher participation. While sharing a name, the OmniScientist detailed in the recent arXiv paper focuses on direct processing of heterogeneous raw evidence, differentiating it from previous work that simulated social scientific infrastructure.
The OmniScientist system's ability to handle diverse data modalities and manage the end-to-end research process from raw evidence to a manuscript represents a significant step in the automation of scientific discovery. The findings indicate that direct perception of raw, multimodal data is a key component for developing more capable AI scientists.