From Solving Problems to Creating Them: A Framework for Discovery Intelligence
Foundation models have gotten very good at solving problems humans hand them. They write code, answer questions, design molecules, and run long workflows. But almost all of this work assumes the problem is already well-defined: the variables are chosen, the objective is fixed, the evaluator determines success. Yang, Yin, and Wu from PHAI Labs argue that the next step is a model system that participates in the process of creating new problems, representations, explanations, and knowledge. They call this capability Discovery Intelligence and formulate Discovery Foundation Models (DFMs) as general-purpose model systems for open-ended discovery.
The distinction is not about model size or reasoning depth. It is about what parts of the research process are fixed inputs and what parts can become objects of model action and revision. A model that searches a predefined design space more efficiently is still operating inside a structure someone else chose. A DFM must be able to construct, test, and revise the structure itself: deciding what is actually unknown, how the problem should be posed, which representation makes competing mechanisms expressible, and what intervention could force them to disagree.
Why Generalist Problem Solving Has a Structural Ceiling
The paper identifies a specific limitation in current foundation model systems. They receive a recognizable object (a question, dataset, benchmark, formal language, design space, or goal) and optimize within it. Search can explore enormous candidate spaces, reinforcement learning can discover unexpected strategies, and agents can automate long workflows. None of these mechanisms guarantees that the supplied variables or evaluator are scientifically adequate.
A fixed representation cannot express a variable it omits. A search objective cannot recover a property that its evaluator systematically ignores. Increasing sample count does not distinguish mechanisms when the observable is non-identifying. An automated workflow can therefore pursue a misframed question more efficiently without becoming better at recognizing the misframing. This structural ceiling becomes visible when a research program stalls: two mechanisms remain observationally equivalent because the measurement is wrong, not the hypothesis. Performance gains disappear under another data split because the evaluator is mismatched, not the optimizer. Every explanation requires local exceptions in the same regime because the representation omits a key variable.
The authors argue that science is the right environment for developing this capability because it exposes these structural decisions. Questions can be underspecified, variables hidden, mechanisms observationally equivalent, interventions costly, evaluators incomplete, and outcomes delayed. Evidence arrives from environments the model cannot rewrite after seeing the result. These properties turn formulation, representation, experiment design, failure attribution, and revision into observable decisions rather than rhetorical qualities of a final answer.
Three Missing Transitions
The gap between predefined problem solving and Discovery Intelligence can be localized to three transitions.
Framing requires moving from observations and uncertainty to a research opportunity worth pursuing. This includes distinguishing persistent structure from noise, deciding which unknowns are consequential and testable, and specifying the scope, scale, conditions, and observables needed to make the problem researchable. A supplied question can be rejected or reformulated when it is too broad, proxy-driven, or impossible to identify under the available measurements.
Modeling requires constructing the variables and abstractions through which explanations become expressible. A useful operation may add a latent variable, remove a proxy, change scale, separate regimes, revise an ontology, or transform the problem into a causal, geometric, symbolic, or programmatic form. Hypotheses are then formed inside this provisional representation and must differ in mechanism, validity conditions, or intervention response rather than only in wording.
Grounding and revision require choosing evidence that can change the status of current explanations and then updating the appropriate research object. A contradiction can indicate theory failure, measurement error, protocol deviation, hidden confounding, simulator misspecification, or environmental shift. Discovery therefore requires both informative intervention and failure attribution. The resulting experience becomes a transferable Discovery Skill only after its trigger and effect survive validation beyond the episode in which it was observed.
Seven Coupled Capabilities
The DFM target is factored into seven coupled capabilities. Problem discovery (C-find) selects unresolved structures worth allocating research resources to and rejects apparent unknowns that disappear under calibration, retrieval, or stronger baselines. Problem formulation (C-form) turns a selected unknown into a bounded and testable problem by fixing its object, scope, scale, conditions, and observables while keeping those choices revisable. Representation construction (C-repr) builds the variables, relations, abstractions, and scales through which the problem is expressed, becoming decisive when the current representation makes every candidate explanation equivalent or repeatedly produces the same failure boundary.
Hypothesis formation (C-hyp) creates mechanistically distinct explanations with explicit assumptions, validity ranges, predictions, and possible falsifiers. Intervention (C-int) chooses experiments, simulations, code executions, ablations, counterexamples, alternative measurements, or replications for their expected effect on the research state rather than for confirmation alone. Evidence-grounded revision (C-rev) attributes unexpected outcomes and updates the appropriate object: hypothesis, representation, problem formulation, protocol, measurement process, or intervention plan. Continual discovery improvement (C-cont) changes future discovery behavior using validated cross-episode experience, specifying when an operation applies, what it should change, and what later evidence would show that the change was beneficial.
The first six capabilities operate within an investigation. The seventh is a cross-task update mechanism. This is stronger than fact accumulation or retrieving a successful trajectory. A reusable operation must survive validation beyond the episode in which it was observed.
Zetema: The System Instantiation
Zetema is the concrete system that instantiates the DFM framework. It maintains an explicit research state that can be revised, branched, and rolled back. It supports verification and experimental gating: consequential actions require validation through a Research World Model before they affect the research state. It connects Dry-Lab reasoning to external computational or physical evidence. And it converts validated cross-task experience into Discovery Skills.
The research state is the central data structure. It records not just what is currently believed but how that belief was formed, what alternatives were considered, what evidence supported or undermined each, and what interventions remain informative. This makes the investigation auditable and reversible. When a contradiction arises, the system can trace it to the specific state transition that introduced the problematic assumption, rather than starting over.
Discovery Skill memory stores validated operations that transfer across tasks. Each skill specifies its trigger conditions, the change it makes to the research state, and what evidence would confirm that the change was beneficial. This is not a trajectory replay mechanism. It is a structured record of when a particular discovery operation worked and why, allowing the system to apply similar operations in new contexts with different domain specifics.
The Research World Model gates consequential actions. Before an intervention is executed, the system checks whether the predicted outcome is consistent with the current research state and whether the intervention is informative under the current hypotheses. This prevents the system from executing experiments that cannot distinguish between current explanations or from updating beliefs in ways that are inconsistent with established evidence.
GALILEO: A Real Discovery Loop with Physical Feedback
The paper grounds the framework with GALILEO, a therapeutic-discovery system that couples Dry-Lab reasoning with robotic and hands-on Wet-Lab experimentation. The system performs multi-omics-informed target nomination and peptide design, robotic synthesis, multimodal phenotyping, orthogonal hands-on assays, and repeated evidence-driven revision. Across experimentally validated LRRC8C and SLC25A1 branches, physical measurements alter subsequent target beliefs, assay choices, mechanism hypotheses, and molecular-design policies.
Across five optimization rounds, the feedback from physical experiments is consolidated into a transferable Amphiphilic Balance Grammar. This is not a final answer. It is a design rule distilled from repeated interaction with a physical system, capturing what worked, when, and why. The grammar encodes the conditions under which a particular molecular design strategy is expected to succeed, based on validated biological feedback rather than computational prediction alone.
The GALILEO case demonstrates the intervention-evidence-revision loop in a physical system. Physical measurements do not merely confirm or disconfirm hypotheses. They alter the research state: changing which targets are prioritized, which assays are chosen, which mechanisms are hypothesized, and which design policies are adopted. The resulting feedback is distilled across rounds into a reusable design rule that specifies its own validity conditions.
The authors keep the broader DFM claim distinct from this domain-specific system. GALILEO shows that the DFM operators (problem discovery, formulation, representation construction, hypothesis formation, intervention, evidence-grounded revision) can form a closed physical discovery loop. It does not show that a general-purpose DFM can perform arbitrary discovery tasks. The distinction matters: a system that performs therapeutic discovery is not the same as a system that can discover in any domain.
Training and Evaluation Beyond Final-Answer Performance
The training formulation targets the intermediate decisions that change a research program. Training data comes from trajectories through interactive research environments, not from static question-answer pairs. Learning objectives include state-conditioned discovery imitation, preference learning over research operations, process verification, Research World Model learning, reinforcement learning from scientific feedback, and validated Discovery Skill learning.
The unified training procedure combines these objectives in a single pass. The system learns to imitate successful research decisions, to prefer operations that advance the research state over those that do not, to verify that proposed state transitions are consistent with evidence, to predict the consequences of interventions, to improve from the outcomes of those interventions, and to extract reusable skills from validated experience.
Evaluation measures both externally validated progress in the current episode and improvement in future discovery behavior under matched resources and retrieval controls. This is a process-centered evaluation: it asks not just whether the system arrived at a correct answer but how it got there, whether the research decisions it made were sound, whether its revisions were appropriate, and whether its experience transfers to new tasks. Stage-wise process evaluation examines each capability independently: did the system identify a valuable unknown, formulate it correctly, construct an adequate representation, form distinct hypotheses, design informative interventions, attribute failures correctly, and extract reusable skills?
The evaluation protocol includes resource-matched comparisons. Two systems are compared under the same computational budget, the same retrieval access, and the same validation mechanisms. This prevents the common failure mode where a system appears to perform well because it received more resources or better tools, not because its discovery decisions were superior.
Four Research Horizons
The paper analyzes how grounding and responsibility change as the framework moves through four regimes. Digital discovery operates over code, databases, and formal systems where feedback is immediate and interventions are cheap. Simulation-grounded discovery adds physical models but retains the ability to run unlimited experiments. Embodied and physical discovery requires robotic or hands-on experimentation where feedback is delayed, experiments are costly, and failures have real consequences. Recursive discovery systems improve their own discovery process, creating a feedback loop where the system's capabilities change its own training distribution.
Each regime imposes different constraints on the DFM operators. Digital discovery allows rapid iteration but risks optimizing against a simulator that does not match reality. Physical discovery enforces real-world grounding but limits the number of experiments. Recursive discovery creates the possibility of self-improvement but also the risk of self-reinforcing errors. The paper argues that the same framework applies across all four regimes, but the governance requirements escalate as the consequences of errors increase.
Epistemic Boundaries and Recursive Risk
The discussion section addresses what the system cannot and should not do independently. Scientific claims require independent validation. The system should preserve which validator supported which state transition and abstain or escalate when the available evidence does not justify promotion of a claim. Provenance and auditability are not optional features: every state transition should record its evidence, its alternatives, and its reasoning.
Scoped authority means the system operates under constraints appropriate to the risk level of its actions. Low-risk actions (reading papers, running existing code) can proceed autonomously. High-risk actions (physical experiments, clinical decisions) require human approval. The autonomy level is orthogonal to the capability level: a system can be highly capable but operate under strict human oversight.
Recursive update risk is the most concerning. When a system improves its own discovery process, it changes the distribution of problems it encounters, the evidence it collects, and the skills it extracts. If the improvement is based on flawed evidence or a misaligned evaluator, the system can reinforce its own errors. The paper argues that recursive improvement requires external validation at every step, not just at the end.
What This Means in Practice
The DFM framework suggests a shift in how we build and evaluate AI systems for science. Instead of measuring how well a model answers a fixed question, we should measure how well it identifies valuable unknowns, formulates testable problems, constructs adequate representations, designs informative interventions, attributes failures correctly, and extracts reusable skills. Instead of training on question-answer pairs, we should train on research trajectories through interactive environments where the system's decisions have observable consequences.
The GALILEO case shows that this is not purely theoretical. A system can couple dry-lab reasoning with physical experimentation, use physical feedback to revise its research state, and distill validated experience into reusable design rules. The five optimization rounds that produced the Amphiphilic Balance Grammar demonstrate that the intervention-evidence-revision loop can produce transferable knowledge, not just task-specific solutions.
For developers, the practical implication is that the most valuable capability of an AI system for science may not be its ability to solve a given problem, but its ability to recognize when the problem is wrong and to construct a better one. This is a harder capability to build and evaluate, but it is the one that determines whether AI participates in the creation of new knowledge or merely optimizes within existing structures.
Read the paper on arXiv