Discovery Foundation Models push AI to discover, not just solve
A paper argues that foundation models have progressed from learning and reasoning over existing knowledge to learning through action, tool use and outcome feedback, and that the next frontier is a further shift: from solving and acting within problems that humans specify to participating in the process by which new problems, representations, explanations and knowledge are created. The authors call this capability Discovery Intelligence.
They formulate Discovery Foundation Models, or DFMs, as general-purpose model systems built for open-ended discovery. A DFM operates over a revisable research state and supports seven coupled capabilities: problem discovery, problem formulation, representation construction, hypothesis formation, intervention, evidence-grounded revision, and continual discovery improvement.
The authors instantiate the framework as a system called 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. GALILEO is presented as the system this discovery loop runs in, not as a system that has already produced a therapeutic result.
The authors then propose a unified approach to capability formation and process-centered evaluation, meant to let discovery behavior be trained, improved and measured beyond final-answer performance. They frame the whole effort as part of a broader progression in intelligence scaling: from learning over existing knowledge, to learning from action outcomes, to participating in building, testing and revising the structures through which new knowledge is discovered. Code for the framework is published on GitHub at Gen-Verse/DFM-Plans.
Key facts
- Discovery Foundation Models (DFMs) are proposed as general-purpose model systems operating over a revisable research state, supporting seven coupled capabilities from problem discovery to continual discovery improvement.
- Zetema instantiates the framework, combining explicit research-state dynamics, verification and experimental gating, external grounding, and cross-task Discovery Skill evolution.
- GALILEO grounds the framework in a real therapeutic-discovery system that closes a physical loop between Dry-Lab reasoning and robotic or hands-on Wet-Lab experimentation, guided by external biological evidence.
- The paper proposes evaluating discovery behavior by process rather than only final-answer accuracy, but the abstract reports no benchmark results or experimental outcomes.
- Code for the framework is released on GitHub under Gen-Verse/DFM-Plans.
Why it matters
Most foundation models today solve problems that a person has already framed. This paper argues the next capability jump is models that help frame the problem, the hypotheses and the representations in the first place, a shift the authors label Discovery Intelligence. If the framework holds up, it changes what counts as capable for a model system: not just answering well, but taking part in the process that produces new questions and new knowledge.
Who it affects
AI researchers building agentic or tool-using systems, and specifically teams working on AI for scientific discovery, drug discovery and lab automation. GALILEO is grounded in a real therapeutic-discovery setting that links Dry-Lab reasoning to physical, robotic and hands-on Wet-Lab experiments, so the framework speaks directly to that audience.
How to use it
The authors publish code for the DFM framework on GitHub at Gen-Verse/DFM-Plans. The abstract gives no pricing, licence terms, or step-by-step description of adopting Zetema or GALILEO beyond the conceptual architecture, so evaluating the work in practice means going to the repository and the full paper for implementation detail.
How solid is it
The abstract lays out a conceptual framework and two instantiations, Zetema and GALILEO, plus a proposed process-centered evaluation approach, but it states no benchmark results, accuracy figures or experimental outcomes. GALILEO is described as the system the discovery loop runs in, not as a system that has already delivered a therapeutic finding.
Risks and caveats
The abstract names no authors or institutional affiliations, gives no publication date or venue, and does not spell out what evidence-grounded revision or Discovery Skill evolution concretely involve beyond naming them. Without reported results, how well the seven capabilities perform in practice cannot be verified from this text alone.