Discovery Foundation Models: Toward Open-Ended Discovery Intelligence
Ling Yang, Zhenfei Yin, Yingcheng Wu, Yingcheng Wu
cs.CL
Quick verdict
An ambitious framework paper defining Discovery Foundation Models; the taxonomy and process protocol are useful, but nothing is implemented, measured, or shown to transfer.
Summary
This position paper argues that foundation models must move beyond solving tasks whose research structure is already fixed, and proposes Discovery Foundation Models as a model-system category for open-ended discovery. The core distinction is formalized by contrasting a conventional task Q = (P, R, G, T, V), where problem, representation, objective, tools and evaluator are supplied, with a discovery episode that maps a partially understood world, an initial knowledge state and constraints to knowledge progress plus an evidence-bearing record of state transitions. Seven coupled capabilities are specified: finding valuable unknowns, formulating researchable problems, constructing representations, forming mechanistically distinct hypotheses, designing discriminating interventions, evidence-grounded revision, and cross-task improvement. The framework is instantiated as Zetema, which maintains an explicit research state with branching and rollback, screens actions through a Research World Model and multi-criterion gating, separates Dry-Lab prediction from Wet-Lab evidence, and promotes Discovery Skills only after transfer validation. Training is specified as a compositional objective over seven losses, and evaluation as two axes, externally validated progress and future capability, with resource-matched controls. Empirical grounding is a single therapeutic case, GALILEO, reporting two organoid IC50 values and five optimization rounds consolidated into an Amphiphilic Balance Grammar.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Current foundation models are evaluated inside research structures that humans already fixed; the missing capability is constructing, testing and revising that structure itself.
- Discovery Intelligence can be factored into seven coupled capabilities spanning problem discovery, formulation, representation, hypothesis formation, intervention, evidence-grounded revision and cross-task improvement.
- A DFM is a model system D = (pi, S, M, T, E, V, H) evaluated as a whole, so autonomy level is orthogonal to capability and human contributions must remain in the provenance record.
- Discovery behaviour can be trained from research-state transitions with a compositional objective, and evaluated on two axes: externally validated progress and improvement in future capability under matched resources.
- The GALILEO therapeutic-discovery loop shows physical evidence revising target beliefs, assay choices, mechanism hypotheses and design policy, and consolidating into a transferable Amphiphilic Balance Grammar.
Strengths
The task-versus-discovery distinction is formalized sharply enough to be argued with
Equation 1 defines a conventional task as Q = (P, R, G, T, V), all supplied, while Equation 2 defines discovery as mapping (W, K0, B) to (delta-K, Xi) where the problem, representation, hypothesis space, intervention strategy and validation procedure may all change mid-episode. Section 2.1 makes the consequence concrete and non-obvious: a fixed representation cannot express a variable it omits, a search objective cannot recover a property its evaluator ignores, and larger sample counts cannot separate mechanisms when the observable is non-identifying.
Negative and non-positive outcomes are first-class, which is rare in AI-for-science writing
Definition 1 explicitly allows a DFM to make progress by showing that an effect does not replicate, that a question is untestable under current measurements, or that a representation omits the variable needed for intervention. Section 5.1 adds explicit branch termination when no feasible intervention is identifying, and Section 4.4 requires the record to retain failed formulations, non-identifying interventions and attribution errors rather than only successes.
The attribution and system-boundary rules are unusually disciplined about credit
The evaluation protocol is designed against the failure modes it names
Section 9 on recursive risk identifies the failure mode that matters most and is genuinely unsettling
Weaknesses
Nothing is implemented, run, or measured; the paper is a specification and a research agenda
The single empirical anchor is second-hand, single-domain, and reported almost without numbers
Every quantity that would make the framework evaluable is left unspecified
The capability definitions are graded and therefore hard to falsify
The novelty claim over adjacent work is asserted rather than demonstrated
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a framework and agenda paper, and it should be read as one. The scope is deliberately large: define a capability category, specify its operators, instantiate one system organization, and propose how to train and evaluate it. The scope discipline is better than most papers of this kind. Section 3.3 insists the evaluated unit is the declared model system D = (pi, S, M, T, E, V, H), that autonomy is orthogonal to capability, and that human contributions stay in the provenance record. Section 5.7 explicitly refuses to let the GALILEO case stand in for the general claim, stating that its domain, modality, interfaces and objectives remain substantially structured and that cross-domain transfer is still open. Five claims carry the argument: the task-versus-discovery distinction; the seven-capability factorization; Zetema's state-centered organization; the two-axis evaluation protocol; and the GALILEO grounding.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is a Discovery Foundation Model?
A Discovery Foundation Model is a proposed category of AI system that does not just answer questions inside a research setup someone else defined, but helps build and revise that setup. The paper contrasts a normal AI task, written as Q = (P, R, G, T, V) where the problem, representation, goal, tools and evaluator are all given, with discovery, where the system starts from a partially understood world and must decide what is worth investigating, how to pose it, which variables make competing explanations distinguishable, and what experiment would force them to disagree. Seven coupled capabilities are listed: finding unknowns, formulation, representation, hypotheses, intervention, evidence-based revision, and improvement across tasks.
How is this different from existing AI scientist systems?
The authors' distinction is about which parts of the research process are fixed inputs and which the system can change. In their comparison table, foundation models for science, scientific reasoning models, scientific agents, search-and-evaluator systems and autonomous laboratories typically receive the problem, or treat representation change and problem revision as optional rather than required. A DFM would require all of: constructing the problem, changing the representation when needed, designing interventions for their informational value, revising the right object when evidence disagrees, grounding claims in external evidence, and transferring validated experience across tasks. The paper notes this is a matter of degree and does not test the classification on any existing system.
Has anything in the paper actually been built and tested?
No. Zetema is specified as functional interfaces: a research state with branching and rollback, a world model and gating layer for consequential actions, separation of computational prediction from physical evidence, and memory updates gated by transfer validation. The training objective combines seven loss terms and the evaluation protocol defines two axes of progress, but no implementation, ablation or baseline is reported. The only empirical content is a second-hand case study, GALILEO, a therapeutic peptide discovery loop reporting two organoid IC50 values (18.922 and 0.736 micromolar) and five optimization rounds that were consolidated into a design rule called the Amphiphilic Balance Grammar.
What risks does the paper itself identify?
The paper devotes a full section to epistemic and governance limits. It argues that novelty is not validity, that model confidence is not scientific certainty, and that independence is not binary: two reviewers can share a base model or two simulations can share assumptions, so reports should state which axes differ. It proposes attaching permission to actions rather than to a global autonomy score, moving from observation and proposal through simulation, approval and execution. The risk it treats as most serious is recursive: if a system can revise its own skills, evaluators or benchmark generators, small distortions can compound and measured performance can improve while genuine discovery capability narrows.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1 through 9 are diagrams and cannot be inspected from the extracted text; all statements come from the body text, equations, Table 1 and figure captions.
- The GALILEO case study is cited to Jiang et al. 2026, which is not included in this document, so its two IC50 values could not be checked against a protocol, sample size or replicate structure.
- Only a plans repository (github.com/Gen-Verse/DFM-Plans) and a collaboration contact appear in the header; no implementation of Zetema, no training data and no benchmark are released, so availability claims are limited to what the header states.
- The paper is a self-described formulation and research agenda; judgments about empirical support should be read as assessments of what the text claims and grounds, not of a working system.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.