Discovery Foundation Models: Toward Open-Ended Discovery Intelligence

Forum topic · 小凯 · 2026-09-16

Summary

A paper posted on zhichai.net introduces Discovery Foundation Models (DFMs), a proposed framework for open-ended scientific discovery by general-purpose AI systems. Authored by Ling Yang, Zhenfei Yin, and Yingcheng Wu (arXiv:2609.15973), the work argues that foundation models should progress beyond learning existing knowledge and acting on human-specified problems toward Discovery Intelligence: creating new problems, representations, explanations, and knowledge. 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. The authors instantiate the framework with Zetema, which couples explicit research-state dynamics, verification and experiment gating, external grounding, and cross-task discovery skill evolution. They ground it in GALILEO, a real therapy-discovery system linking in-dry-lab reasoning, robotic wet-lab experiments, external biological evidence, and iterative hypothesis revision in a closed physical discovery loop. A unified capability-formation, process-centric evaluation approach makes discovery trainable, improvable, and measurable beyond final-answer performance.

Paper Overview

Field: NLP
Authors: Ling Yang, Zhenfei Yin, Yingcheng Wu
Posted: 2026-09-14
arXiv: 2609.15973

Abstract

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.

Key Ideas

Discovery Foundation Models (DFMs) are formulated as general-purpose model systems for open-ended discovery.
A DFM operates over a revisable research state and supports seven coupled capabilities:
Zetema instantiates the framework, coupling explicit research-state dynamics, verification and experiment gating, external grounding, and cross-task discovery skill evolution.
GALILEO, a real therapy-discovery system, grounds the framework: in-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.
A unified capability-formation and process-centric evaluation approach is proposed so that discovery behavior can be trained, improved, and measured beyond final-answer performance.

Together, these components establish discovery as a learnable, executable, and evaluable capability of foundation model systems — part of a broader trajectory of intelligence: from learning existing knowledge, to learning from action outcomes, to constructing the structures of discovering, testing, and revising new knowledge.

---

*Auto-collected on 2026-09-16*