AIAny
Icon for item

OmniScientist: An Omni-Modal Omni-Discipline AI Scientist

Conducts end-to-end multidisciplinary research directly from heterogeneous raw evidence using lifecycle-wide perception and three autonomous agents (Ideation, Experiment, Writeup). Integrates perceptual analysis, execution provenance, and code-enforced checks to produce executable analyses, validated results, and compiled manuscripts across many modalities.

Introduction

Most automated research systems limit themselves to text, labels, or precomputed summaries; that leaves spatial, temporal, cross-channel and procedural relations—often decisive in scientific work—out of the loop. OmniScientist's core insight is that keeping raw observation and perception available throughout the research lifecycle changes what an AI scientist can ask, test, and claim.

Key Findings
  • Lifecycle design: a perception layer plus three sequential autonomous agents (Ideation → Experiment → Writeup) run inside a deterministic pipeline so observations can reshape hypotheses, experimental choices, and final claims. This enforces numerical traceability and execution provenance via code checks.
  • Empirical evaluation: completed full raw-data→paper workflows on 36 real-data cases spanning 5 discipline families and 12 modalities (images, signals, audio, video, 3-D structures, trajectories, tables, formulae, graphs). Generated papers scored a mean overall paper score of 6.3 (reference backbone).
  • Value of perception: compared to a blind baseline that only saw scalar features, direct perception improved all evaluated dimensions and won 85% of head-to-head judgments, indicating perceptual access materially changes question selection, analyses, and supported claims.
Who it's for and tradeoffs

Great fit if you need an evidence-grounded research pipeline that must reason over spatial/temporal/multi-channel data and produce reproducible execution records and manuscripts. It is particularly relevant for multidisciplinary teams prototyping automated analysis across heterogeneous modalities. Look elsewhere if you need a lightweight prompt-only pipeline (OmniScientist prioritizes perceptual tooling, execution tracing, and code-enforced rigour, which adds system complexity and compute requirements).

Mechanisms (brief)

The system groups evidence into four discipline-independent families (perceptual, symbolic, quantitative-statistical, procedural), exposes perception tools (spatial, temporal, cross-channel, statistical, dynamic), and performs idea/rigour/claim checks in code (novelty screening, statistical validity, anti-HARKing, multiple comparisons, provenance). The result is an end-to-end flow from raw artifacts to executable experiments and a compiled manuscript.

Information

  • Websitearxiv.org
  • Organizations1National University of Singapore 2University of Oxford, Project page: https://omni-scientist.github.io, Software & Skill: https://github.com/Omni-Scientist/OmniScientist
  • AuthorsBobo Li, Hao Fei, Tianjie Ju, Mong-Li Lee, Wynne Hsu
  • Published date2026/08/13

Categories

More Items

Blocks tool-using agents from directing emails, payments, and other side effects to unauthorized targets by checking typed tool-return fields. Deterministic rules also remove instruction-like text before the model reads it.

Provides six open language models from 0.9B to 375B-A23B for reasoning and agentic tasks. The paper also introduces sparse attention and a decoding speedup that needs no separate draft model.

Proposes a framework for robots that combine foundation models with an embodied runtime to interpret intent, call skills, verify outcomes, and adapt. It also outlines an evaluation framework and a roadmap toward sustained autonomy.