Virtuous Machines: Towards Artificial General Science

TL;DR

The paper addresses the bottleneck in scientific progress by building a domain-agnostic, agentic AI Scientist that can autonomously navigate the full scientific workflow from hypothesis generation to manuscript preparation. It implements a hierarchical multi-agent system with cognitive operators and dynamic memory (d-RAG), leveraging a mixture of frontier LLMs to ideate, design, execute online studies, analyze data, and compose reports. In cognitive psychology, the system autonomously conducted three studies with 288 online participants, producing three complete manuscripts in ~17 hours per study, with expert review confirming rigorous methods but noting conceptual nuances and interpretation gaps. The work demonstrates a step toward embodied AI that can actively test hypotheses in real-world settings and raises important questions about scientific understanding, credit, safety, and governance as AI-driven discovery scales across domains.

Abstract

Artificial intelligence systems are transforming scientific discovery by accelerating specific research tasks, from protein structure prediction to materials design, yet remain confined to narrow domains requiring substantial human oversight. The exponential growth of scientific literature and increasing domain specialisation constrain researchers' capacity to synthesise knowledge across disciplines and develop unifying theories, motivating exploration of more general-purpose AI systems for science. Here we show that a domain-agnostic, agentic AI Scientist system can independently navigate the scientific workflow - from hypothesis generation through data collection to manuscript preparation. The system autonomously designed and executed three psychological studies on visual working memory, mental rotation, and imagery vividness, executed one new online data collection with 288 participants, developed analysis pipelines through 8-hour+ continuous coding sessions, and produced completed manuscripts. The results demonstrate the capability of AI scientific discovery pipelines to conduct non-trivial research with theoretical reasoning and methodological rigour comparable to experienced researchers, though with limitations in conceptual nuance and theoretical interpretation. This is a step toward embodied AI that can test hypotheses through real-world experiments, accelerating discovery by autonomously exploring regions of scientific space that human cognitive and resource constraints might otherwise leave unexplored. It raises important questions about the nature of scientific understanding and the attribution of scientific credit.

Figures (5)

• Figure 1: $|$ Simplified network architecture of the autonomous scientific discovery system. Directed graph illustrating the information flow and functional relationships between agents. The master agent (purple) coordinates the core scientific workflow agents (green) including method, data analysis, and visuals. Expansion to subagent modules (shades of brown) provides domain-specific capabilities, and interactions with specialist agent pathways (blue) handle specialised tasks including coding, review, troubleshooting, and inspection processes. The stacked panel agent boxes indicate agents completing tasks in parallel, while the dotted connections to blank boxes represent further agent lineages in the system not shown here for clarity. Arrows indicate bidirectional data flow between modules and across hierarchical levels. The distributed architecture enables offloading of complex research tasks while maintaining coherent experimental narratives through the centralised master coordinator.
• Figure 2: $|$ Hierarchical framework of cognitive agency levels. Concentric layers represent ascending levels of agent sophistication, from basic retrieval mechanisms to advanced metacognitive capabilities. Each level encompasses the functionalities of those beneath it while introducing emergent properties. Retrieval forms the foundational layer, providing access to external information and internal memory stores. Abstraction enables pattern recognition and generalisation beyond specific instances. Metacognition introduces self-monitoring and strategic control of cognitive processes. Decomposition allows complex problems to be systematically partitioned into manageable components. Autonomy confers goal-directed behaviour independent of external guidance. Collaboration represents the highest level, enabling coordinated multi-agent interaction and collective problem-solving.
• Figure 3: $|$ Iterative experimentation cycles allow the system to extend beyond trained-on knowledge. Schematic representation of the self-directed research process executed by the autonomous discovery pipeline. Each circle represents a complete experimental iteration, with transitions between cycles driven by hypothesis refinement and knowledge accumulation. The Internal/External Seed initiates the research trajectory, establishing initial hypotheses or responding to external queries. Progressive cycles demonstrate the capacity for autonomous experimental design, execution, and interpretation, with each iteration informed by previous outcomes. The expanding experimental space explored through successive iterations is characteristic of open-ended scientific discovery.
• Figure 4: $|$ Three-phase ideation process for hypothesis generation. Schematic representation of the iterative scientific ideation workflow implemented to develop novel research questions. Phase 1 (Generation) produces unique, ranked idea suggestions by brainstorming initial concepts, removing redundancies, filtering impractical proposals, ranking by scientific merit, and re-ranking to refine prioritisation. Phase 2 (Formulation) develops the suggestions into detailed research ideas through expansion of conceptual scope, formulation of testable hypotheses, review to interrogate validity, and iterative improvement. Phase 3 (Validation) yields final vetted research ideas via literature review for context, novelty checking against existing work, feasibility assessment for methodological adaptability, and finalisation of a structured research proposal
• Figure 5: $|$ Manuscript generated by the pipeline. The full 31-page manuscript produced in Study 1 spans hypothesis formulation to final formatting and follows a standard scientific manuscript structure with embedded figures (visible in pages 6, 7, 11, 13, 15, 17, 28, 30), tables (pages 12 & 27), statistical outputs, and references. All content was autonomously generated and typeset in LaTeX. This represents one of the three independently generated manuscripts.