A newly posted paper on Hugging Face, titled Discovery Foundation Models: Toward Open-Ended Discovery Intelligence, argues that foundation models are moving through successive stages: from learning and reasoning over existing knowledge, to learning through action, tool use, and outcome feedback. The authors contend 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.
The paper names this capability Discovery Intelligence and formulates Discovery Foundation Models (DFMs) as general-purpose model systems for open-ended discovery. According to the authors, a DFM operates over a revisable research state and supports seven coupled capabilities: problem discovery, formulation, representation construction, hypothesis formation, intervention, evidence-grounded revision, and continual discovery improvement.
To instantiate the framework, the authors present Zetema, which couples explicit research-state dynamics, verification and experimental gating, external grounding, and cross-task Discovery Skill evolution. They further ground the framework with GALILEO, described as 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.
The paper also formulates a unified approach to capability formation and process-centered evaluation, aiming to enable discovery behavior to be trained, improved, and measured beyond final-answer performance. The authors conclude that these components together establish discovery as a learnable, executable, and evaluable capability of foundation-model systems, framing the 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.
Code associated with the work is available at github.com/Gen-Verse/DFM-Plans.