Discovery Foundation Models: Toward Open-Ended Discovery Intelligence

AI summary

Overview

Research area: AI foundation models, autonomous scientific discovery, and AI-for-Science system design (cross-cutting NLP, machine learning, robotics, and laboratory automation).

Technical level: Intermediate. The paper is primarily conceptual and framework-driven, supported by formal definitions and equations, but it does not require deep mathematical background to follow the central argument.

Scope in one sentence: The paper proposes "Discovery Foundation Models" as a new capability category for AI systems that do not just solve human-specified problems but participate in constructing, testing, and revising the research structures through which new knowledge is created, and it offers one system design (Zetema) plus one real therapeutic-discovery case study (GALILEO) as grounding.

What This Paper Is About

Current foundation models are strong at solving problems whose structure is already fixed by humans: the question is stated, the variables are supplied, the objective is set, and an evaluator decides if the answer is acceptable. The paper argues that this interface hides a ceiling, because many real scientific bottlenecks require changing the research structure itself—reformulating the question, introducing a missing variable, replacing a proxy, or designing an intervention that separates competing explanations. The goal is to define, instantiate, and evaluate a model-system category capable of open-ended discovery rather than optimization inside a predefined task.

Key Contributions

1. Formulates Discovery Foundation Models (DFMs) as a capability-based model-system category, and specifies the research objects and operations that distinguish open-ended discovery from optimization over a predefined task. It defines three required properties: transfer beyond a single fixed task, evidence-appropriate external grounding (not model confidence), and evaluation of the declared model system including state, memory, tools, environments, validators, and human participation.

2. Defines the Discovery Process and instantiates it with Zetema, a system that couples explicit research-state revision, evidence-based action gating through a Research World Model, external grounding (Dry-Lab and Wet-Lab interfaces), and cross-task "Discovery Skill" evolution with attribution.

3. Formulates training and evaluation mechanisms for learning discovery operations, allocating resources across the process, and measuring externally validated knowledge progress plus transferable improvement on unseen tasks under matched resources and retrieval controls.

4. Empirically grounds the Dry-Lab/Wet-Lab loop with GALILEO, a real therapeutic-discovery system in which physical biological feedback revises subsequent scientific decisions and is distilled across rounds into a reusable design rule.

Main Findings

• The gap is structural, not just knowledge or reasoning: Most training and evaluation pipelines reward competence after the research structure is fixed. Fixed representations cannot express omitted variables, search objectives cannot recover properties their evaluators ignore, and more samples cannot distinguish mechanisms when the observable is non-identifying.

• Three missing transitions define the target: Framing (turning observations into a researchable problem), Modeling (constructing variables, abstractions, and representations), and Grounding and Revision (choosing informative evidence and correctly attributing failures to theory, measurement, protocol, confounding, or simulator misspecification).

• Seven coupled capabilities constitute a DFM: Finding valuable unknowns (C_find), formulating researchable problems (C_form), constructing and revising representations (C_repr), forming mechanistically distinct hypotheses (C_hyp), designing interventions (C_int), evidence-grounded revision (C_rev), and continual cross-task improvement (C_cont).

• Discovery can succeed with negative results: Showing an effect does not replicate, a question is untestable under current measurements, or a representation omits a required variable is a valid DFM output when supported by the research state.

• Zetema maintains an explicit, typed research state: The state stores active problem formulations, representations, hypotheses, evidence with provenance, interventions, budgets, and memory, with branches treated as first-class objects that can be compared, merged, archived, or restored. Rejected branches are not deleted, and a representation revision is not stored as another hypothesis.

• Discovery Skills are validated procedures, not advice: A skill is stored as a state-dependent trigger, an operation, an expected effect, and the validation evidence required—distinguishing world facts, episodic memory, and genuine procedural improvement. Only the third kind counts as Discovery Skill improvement.

• A Research World Model screens actions before they become expensive: It predicts observations, state changes, costs, and risks to filter fluent but redundant, infeasible, unsafe, or empirically indistinguishable candidates. The paper explicitly notes this creates a selection effect: world-model error changes which research directions get resources.

• Capability comparison across paradigms: Table 1 contrasts DFMs with foundation models for science, scientific reasoning models, scientific agents/AI-scientist systems, search-and-evaluator systems, and autonomous laboratories. DFMs are unique in requiring constructed problems, representation change, intervention design, problem revision, external grounding, and skill transfer together.

• GALILEO grounds the loop physically: Across experimentally validated LRRC8C and SLC25A1 branches, physical measurements altered subsequent target beliefs, assay choices, mechanism hypotheses, and molecular-design policies; across five optimization rounds this feedback consolidated into a transferable "Amphiphilic Balance Grammar."

Methodology in Plain English

The authors take a conceptual and formalization-first approach. They begin by contrasting a conventional model task—abstracted as problem, representation, goal, tools, and evaluator—with a discovery episode that maps a partially understood world, an initial knowledge state, and constraints onto validated knowledge progress plus a record of how the investigation changed. They then define the capability set, specify the system boundary as a tuple of policy, state, memory, tools, environment, validators, and human oversight, and set out the operators of the Discovery Process, including a formalized intervention-selection objective that maximizes expected reduction in uncertainty over competing explanations. To show the framework is operational rather than purely theoretical, they specify Zetema as an architecture-agnostic organization with explicit state dynamics, verification gating, and skill memory. Finally, they connect the digital framework to physical reality through GALILEO, where robotic synthesis, multimodal phenotyping, and hands-on assays return evidence the model cannot rewrite, driving iterative revision of targets, hypotheses, and molecule design rules.

Why This Matters

Impact on research: The paper reframes the evaluation target for AI-for-Science systems—from final-answer performance to the quality of intermediate research decisions, attribution, and transferable discovery behavior. It provides vocabulary and a capability checklist that can be used to audit whether a given "AI scientist" system actually changes research structure or only optimizes inside a human-supplied one. It also argues that humans and validators must be inside the declared system boundary for honest attribution, which has direct implications for how results are reported.

Real-world applications:

• Therapeutic discovery: multi-omics-informed target nomination, peptide design, and iterative optimization, as demonstrated in the GALILEO case.
• Autonomous and self-driving laboratories: connecting proposals to robotic synthesis and physical measurement with evidence-based gating before costly experiments.
• Materials and molecular design: revising design representations and evaluators rather than only searching a fixed design space.
• AI-assisted research workflows: literature-grounded analysis, data analysis, and long-horizon research program management where reformulation and failure attribution matter.

Industry relevance: Pharmaceutical, biotechnology, chemicals, and materials companies that invest in automated experimentation and computational discovery pipelines are the most direct beneficiaries. The framework also matters to teams building scientific agents or AI-scientist products, because it offers a concrete alternative to benchmark-chasing and a way to distinguish genuine capability gains from fluent but weakly grounded output. The paper includes a collaboration program inviting scientists and experimental platforms with real problems and validation conditions.

Future Directions

• Standardizing process-centered evaluation: How to measure discovery behavior and transferable improvement under matched resources and retrieval controls, without collapsing everything into a single scalar reward.
• Training discovery operations at scale: What supervision signals—trajectories, interactive environments, scientific feedback, resource-allocation data—best train formulation, representation revision, and failure attribution, and how these generalize across domains.
• Scaling from digital to physical and recursive settings: The paper flags open questions about how grounding requirements and responsibility shift as the same framework moves from Dry-Lab reasoning to Wet-Lab execution and to recursive discovery loops.
• Validating and constraining Discovery Skills: Determining when a candidate lesson is genuinely reusable rather than a retrospective explanation dependent on privileged information or unrecorded human correction, and how to weaken or retire skills whose expected effects fail to appear.

Target Audience

AI researchers working on foundation models, agents, and AI-for-Science; laboratory automation and self-driving-lab engineers; computational and experimental scientists in drug discovery, biology, chemistry, and materials; research leaders and policymakers evaluating claims about automated scientific discovery; and anyone designing benchmarks or evaluation protocols for long-horizon research systems. Readers looking for a concrete model architecture or experimental benchmark results will find this paper is a framework and position piece rather than an empirical methods paper.

Authors’ 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. We instantiate this framework with Zetema, which couples explicit research-state dynamics, verification and experimental gating, external grounding, and cross-task Discovery Skill evolution. We further ground the framework with GALILEO, 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. We then formulate a unified approach to capability formation and process-centered evaluation, enabling discovery behavior to 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. We view this 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: https://github.com/Gen-Verse/DFM-Plans