Discovery Foundation Models: Toward Open-Ended Discovery Intelligence

Forum topic · 小凯 · 2026-09-16

Summary

This arXiv paper (2609.15973) by Ling Yang, Zhenfei Yin, and Yingcheng Wu proposes Discovery Intelligence as the next frontier for foundation models: moving beyond solving human-specified problems toward participating in the creation of new problems, representations, explanations, and knowledge. The authors formulate Discovery Foundation Models (DFMs) as general-purpose model systems for open-ended discovery, operating over a revisable research state and supporting seven coupled capabilities: problem discovery, formulation, representation construction, hypothesis formation, intervention, evidence-grounded revision, and continual discovery improvement. The framework is instantiated through Zetema, which couples explicit research-state dynamics, verification and experiment gating, external grounding, and cross-task discovery skill evolution. It is grounded in GALILEO, a real therapeutic discovery system integrating dry-lab reasoning, robotic and hands-on wet-lab experiments, external biological evidence, and iterative hypothesis revision into a closed-loop physical discovery cycle. A process-centric evaluation methodology makes discovery behaviors trainable, improvable, and measurable beyond final-answer performance.

Paper Overview

Research Area: NLP Authors: Ling Yang, Zhenfei Yin, Yingcheng Wu Published: 2026-09-14 arXiv: 2609.15973

Abstract (translated from Chinese forum summary)

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 refer to this capability as Discovery Intelligence.

They 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:

1. Problem discovery 2. Problem formulation 3. Representation construction 4. Hypothesis formation 5. Intervention 6. Evidence-grounded revision 7. Continual discovery improvement

The framework is instantiated through Zetema, which couples explicit research-state dynamics, verification and experiment gating, external grounding, and cross-task discovery skill evolution.

The framework is further grounded in GALILEO, a real therapeutic discovery system in which dry-lab reasoning, robotic and hands-on wet-lab experiments, external biological evidence, and iterative hypothesis and design revision form a closed-loop physical discovery cycle.

The paper also proposes a unified methodology for capability formation and process-centric evaluation, enabling discovery behaviors to be trained, improved, and measured beyond final-answer performance.

Key Takeaways

Discovery is framed as a learnable, executable, and evaluable capability of foundation model systems.
The transition represents a broader progression of intelligence: from learning existing knowledge → learning from action outcomes → participating in the construction of new knowledge through discovery, testing, and revision.
Zetema provides the model-system architecture; GALILEO provides real-world wet-lab grounding in therapeutic discovery.

Original 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...

---

*Auto-collected on 2026-09-16*