Discovery Foundation Models: Toward Open-Ended Discovery Intelligence
DFM takes an AI through a whole turn of discovery: find a question worth asking, test it against real evidence, keep what worked for next time.
Move the cursor over the diagram
Overview
A research paradigm around Discovery Intelligence
Turns “discovery” into a capability of model systems that can be learned, executed and evaluated.
DFM (Discovery Foundation Models) is neither a single model nor a software framework that forces every system through one pipeline. It is the research paradigm, capability definition and system organisation that PhAI Labs proposes around Discovery Intelligence: through formal definitions, process diagrams, system instances and real scientific cases, it turns “discovery” from a vague notion into a capability of model systems that can be learned, executed and evaluated.
In this process, the scientific question comes first. Starting from a partially understood world, the system identifies a valuable unknown, turns it into a researchable problem, and builds hypotheses around it. A hypothesis does not stay on paper. The system calls research tools, data and experimental environments to act on it, gathers external evidence from analysis, experiments and expert judgement, and uses that evidence to confirm, revise or reject its own position.
Model training sits inside this process rather than in front of it. Every tool call and every experimental result is both a test of the model’s judgement and a training signal with a real outcome attached. Validated knowledge, reusable questions, tools and verifiers accumulate, and become the starting point for the next discovery.
Around this paradigm and capability definition, the research unfolds through system instances, independent papers, engineering modules and real scientific collaborations, released progressively. They do not add up to one fixed implementation; they explore, from different directions, how Discovery Intelligence becomes operational.
The seven capabilities DFM is about
- Identify valuable unknowns
- Turn unknowns into researchable questions
- Build and revise the representation
- Build and revise hypotheses
- Call tools, data and experimental environments
- Validate and update against external evidence
- Accumulate reusable research structure and discovery capability
The DFM research-map repository carries the conceptual structure of the Discovery Loop, research directions, the open-source plan, demos and how to collaborate. Its directories and direction markers reserve future scope in public; they do not mean the corresponding implementations are finished or released.
Abstract
Foundation models have progressed from learning and reasoning over existing knowledge, to increasingly learning through action, tool use, and outcome feedback. We argue that the next frontier is a further transition: from solving and acting within problems specified by humans to participating in the process by which new problems, representations, explanations, and knowledge are created. We refer to this capability as Discovery Intelligence. We formulate Discovery Foundation Models (DFMs) as general-purpose model systems for open-ended discovery. A DFM operates over a revisable research state and supports seven coupled capabilities spanning problem discovery, formulation, representation construction, hypothesis formation, intervention, evidence-grounded revision, and continual discovery improvement. We instantiate this framework with Zetema, which couples explicit research-state dynamics, verification and experimental gating, external grounding, and cross-task Discovery Skill evolution. We further ground the framework with GALILEO, a real therapeutic-discovery system in which Dry-Lab reasoning, robotic and hands-on Wet-Lab experimentation, external biological evidence, and iterative hypothesis and design revision form a closed physical discovery loop. We then formulate a unified approach to capability formation and process-centered evaluation, enabling discovery behavior to be trained, improved, and measured beyond final-answer performance. Together, these components establish discovery as a learnable, executable, and evaluable capability of foundation-model systems. We view this shift as a broader progression in intelligence scaling: from learning over existing knowledge, to learning from action outcomes, and ultimately to participating in the construction, testing, and revision of the structures through which new knowledge is discovered.
Authors
Venue
Cite
BibTeX
@article{yang2026discovery,
title={Discovery Foundation Models: Toward Open-Ended Discovery Intelligence},
author={Yang, Ling and Yin, Zhenfei and Wu, Yingcheng},
journal={arXiv preprint arXiv:2609.15973},
year={2026}
}