Discovery Foundation Models: Toward Open-Ended Discovery Intelligence

A paper introduces Discovery Foundation Models (DFMs), general-purpose systems for open-ended discovery built on seven coupled capabilities, instantiated in the Zetema framework and the GALILEO therapeutic-discovery system that closes the loop between dry-lab…

Two sources (Hugging Face daily papers, 2026-09-13, and arXiv cs.AI/cs.LG/cs.CL, 2026-09-14) report the same paper, which formulates Discovery Foundation Models (DFMs) as general-purpose model systems for open-ended discovery. Rather than only solving given tasks, DFMs are designed to participate in creating new problems and new knowledge, which the authors frame as 'Discovery Intelligence' — positioned as the next frontier beyond tool-use and reasoning. The framework specifies seven coupled capabilities: problem discovery, problem formulation, representation construction, hypothesis formation, intervention, evidence-grounded revision, and continual improvement. It is instantiated with Zetema, which combines explicit research-state dynamics, verification gating (Report 1 also describes 'experimental gating'), external grounding (per the arXiv report), and cross-task Discovery Skill evolution. The framework is grounded with GALILEO, a real therapeutic-discovery system that combines Dry-Lab reasoning with robotic and hands-on Wet-Lab experimentation in a closed physical discovery loop, including iterative hypothesis revision. Code is released on GitHub. The arXiv report additionally notes the authors define process-centered evaluation, so discovery behavior can be trained and measured beyond final answers. The two reports are consistent; the arXiv version adds detail (external grounding, hands-on wet-lab work, iterative revision, process-centered evaluation).

• Paper title: 'Discovery Foundation Models: Toward Open-Ended Discovery Intelligence'; surfaced on Hugging Face daily papers 2026-09-13 and arXiv (cs.AI/cs.LG/cs.CL) 2026-09-14
• DFMs are general-purpose model systems for open-ended discovery intended to create new problems and knowledge, not just solve given tasks
• Authors frame 'Discovery Intelligence' as the next frontier beyond tool-use and reasoning
• Framework defines seven coupled capabilities: problem discovery, problem formulation, representation construction, hypothesis formation, intervention, evidence-grounded revision, and continual improvement
• Zetema instantiates the framework with explicit research-state dynamics, verification (and experimental) gating, external grounding, and cross-task Discovery Skill evolution
• GALILEO is a real therapeutic-discovery system coupling Dry-Lab reasoning with robotic and hands-on Wet-Lab experimentation in a closed physical discovery loop, including iterative hypothesis revision
• Code for the work is released on GitHub
• The paper also defines process-centered evaluation to train and measure discovery behavior beyond final answers (arXiv report)