Discovery Foundation Models: Toward Open-Ended Discovery Intelligence

Discovery Foundation Models. Toward Open-Ended Discovery Intelligence. By Ling Yang and two others. Published on the fourteenth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Foundation models have progressed from learning and reasoning over existing knowledge to increasingly learning through action, tool use, and outcome feedback. The authors 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. They name this capability Discovery Intelligence. The shift is from answering questions to taking part in deciding which questions deserve asking, and in producing the knowledge that answers them.

The organising object is the Discovery Foundation Model, or DFM, formulated as a general-purpose model system for open-ended discovery. A DFM operates over a research state that can be revised, and it supports seven coupled capabilities. They span problem discovery, formulation, representation construction, hypothesis formation, intervention, evidence-grounded revision, and continual improvement of discovery itself. The seven capabilities trace the arc of a research project, from noticing a problem to improving the way discovery itself is done. Coupled is the operative word: they are meant to work together over a revisable state, not as isolated skills.

Two systems ground the framework. Zetema instantiates it, coupling explicit research-state dynamics, verification and experimental gating, external grounding, and Discovery Skill evolution across tasks. GALILEO is a real therapeutic-discovery system, in which Dry-Lab reasoning, robotic and hands-on Wet-Lab experimentation, external biological evidence, and iterative revision of hypotheses and designs form a closed physical discovery loop. On top of these, the authors formulate a unified approach to capability formation and to process-centered evaluation, so that discovery behavior can be trained, improved, and measured beyond final-answer performance. GALILEO carries the loop into the physical laboratory, where hypotheses are tested by real experiments and then revised.

One line runs through the whole piece: a progression in how intelligence scales. First, learning over existing knowledge. Then, learning from action outcomes. And finally, participating in the construction, testing, and revision of the structures through which new knowledge is discovered. Each stage keeps what came before and adds a further way of learning.

A second line concerns what counts as success. The emphasis falls on the process of discovery, a revisable state, verification before experiments proceed, and evidence that feeds back into revision, rather than on a single final answer. Together these components are meant to establish discovery as a capability of foundation-model systems that is learnable, executable, and evaluable. If discovery is to be trained, it has to be measured, and it cannot be measured by final answers alone.

This is a framework and an argument more than a single result. Its claim is that discovery can be treated like other capabilities a model system has, specified, trained and measured, and it offers Zetema and GALILEO as first instances. The code is public.