OmniScientist: An Omni-Modal Omni-Discipline AI Scientist

Analysis

Why This Paper Matters

OmniScientist addresses a critical gap in AI-driven scientific discovery: existing systems typically reason over text, code, labels, or precomputed summaries, missing the rich spatial, temporal, cross-channel, and procedural information inherent in raw scientific data. By enabling an AI scientist to perceive and reason over heterogeneous raw evidence—such as images, signals, audio, video, 3D structures, trajectories, tables, formulae, and graphs—OmniScientist moves beyond workflow automation to true evidence-grounded discovery. This is a significant step toward AI systems that can autonomously conduct research across disciplines, not just within narrow domains.

The paper's emphasis on lifecycle-wide perception—where observations shape research questions, experimental decisions, and final claims—is a paradigm shift. It contrasts with prior work that treats perception as a preprocessing step, instead integrating it throughout the research pipeline. This holistic approach is likely to inspire future AI scientist frameworks to prioritize raw data access and multimodal reasoning.

Technical Contributions

• Omni-modal perception layer: Encodes raw data from multiple modalities into a unified representation, making scientifically decisive information accessible to the agent.
• Three autonomous agents: Ideation, experiment, and writeup agents operate in a deterministic pipeline, ensuring that each stage is informed by the raw evidence.
• Code-based checks: Idea, rigour, and claim checks are implemented in code, providing novelty screening, statistical validity, execution provenance, and numerical traceability—critical for scientific integrity.
• End-to-end automation: The system completes the full research lifecycle from raw data to a compiled manuscript, demonstrating a practical, reproducible pipeline.
• Evaluation framework: A comprehensive benchmark across 36 real-data cases spanning 5 discipline families and 4 evidence families, with a blind variant for controlled comparison.

Results

OmniScientist achieved a mean overall paper score of 6.3 with the reference reasoning backbone, and completed all 36 cases from raw data to compiled manuscript. In paired comparisons against a blind variant that only receives precomputed scalar features, direct perception improved all 7 evaluation dimensions and won 85% of head-to-head judgments. These results strongly support the hypothesis that perception is essential for evidence-grounded scientific discovery.

The 85% win rate is particularly compelling, as it quantifies the benefit of raw data access over simplified feature representations. The improvement across all evaluation dimensions—likely including novelty, rigour, and clarity—indicates that perception enhances not just data interpretation but the entire research process.

Significance

OmniScientist sets a new benchmark for AI scientists by demonstrating that omni-modal perception is not just a nice-to-have but a necessity for evidence-grounded discovery. This work could accelerate research in fields where data is inherently multimodal, such as biomedical imaging, climate science, and materials science. By automating the full research pipeline with rigorous checks, it also has the potential to improve reproducibility and reduce human bias in scientific workflows.

The broader impact extends to the AI community: it challenges the assumption that precomputed features are sufficient for complex reasoning tasks and encourages the development of models that can directly consume raw data. As foundation models continue to advance, OmniScientist provides a blueprint for building AI scientists that can autonomously explore new frontiers, potentially leading to discoveries that would be difficult for humans alone. However, the reliance on a reference reasoning backbone and the limited number of cases suggest that further scaling and validation are needed before widespread adoption.