Discovery Foundation Models and Autonomous Research

Discovery Foundation Models (DFMs) replace static research pipelines with agentic systems that iterate through problem discovery, formulation, representation construction, hypothesis formation, intervention, and evidence-grounded revision. Traditional foundation models map input prompts to output answers based on static training data. DFMs manage a persistent research state designed to guide scientific inquiry through a sequence of self-directed experiments. This transition shifts the model's role from a pattern-matching engine to a laboratory participant. One operational goal of this architecture is to anchor the model’s reasoning to an accumulated history of experimental results rather than a single prompt.

Managing Research State

Unlike static RAG, this architecture uses persistent state to update hypotheses based on intervention outcomes. The system writes every experimental outcome to a persistent vector store. When the model generates a new hypothesis, it performs a semantic similarity search across the vector store to retrieve the top-k most relevant past outcomes. These retrieved results are injected into the system prompt with a specific instruction to prioritize counter-evidence, forcing the model to explicitly re-evaluate its previous parameters in light of failure data. By requiring the model to draft a 'reconciliation report' comparing new outcomes to the initial hypothesis before committing to a next step, the system prevents the model from ignoring contradictory lab findings. This creates a record of scientific inquiry that acts as the model's current standing, ensuring every new hypothesis is conditioned on the accumulated history of previous attempts and empirical evidence.

Instantiating Discovery

Zetema is a primary system instantiation of this framework, while GALILEO serves as a concrete implementation that bridges wet-lab and dry-lab experimentation to ground the framework in therapeutic discovery. These systems employ automated verification to process experimental results. When an intervention yields data, the system evaluates the finding against the existing evidence record. If the retrieved outcomes are contradictory, the revision cycle is triggered. During this cycle, the model does not merely re-prompt; it re-executes the hypothesis formulation phase while explicitly appending the specific failed experimental data to the prompt context. This forces the model to treat the negative result as a hard constraint that must be resolved before the system can commit a new, revised intervention to the persistent store.

Consider a model attempting to identify a protein binder. It initiates a simulation to predict binding affinity and commits that result to the persistent state. When a physical binding assay fails, the model identifies the discrepancy between its predicted affinity and the real-world lab failure. Because the system retrieves its own history, it must reconcile the failed assay result as a high-priority context item. The model cannot simply ignore the failure because the search logic mandates that the most recent and relevant outcomes define the latent space of the current reasoning session. This forces the model to generate a new hypothesis that accounts for the binding failure, moving the reasoning process forward through empirical contradiction rather than internal logic alone.

Practical Implications for Builders

Building these systems requires confronting the risk of circular reasoning, where the model reinforces nodes without incorporating contradictory evidence. Standard reinforcement learning rewards often optimize for a specific prediction score, which can encourage the model to enter self-consistent, redundant loops that mimic progress. In these scenarios, the vector retrieval mechanism fails to act as a circuit breaker because the model is trained to favor high-confidence, internally consistent outputs that naturally mimic the distribution of successful past experiments. The retrieval mechanism is often biased toward results that match the current hypothesis structure, effectively masking negative data.

Success is instead defined by the model's ability to maintain a high-entropy state, meaning a diversity of potential research paths that remains sensitive to negative empirical data. Maintaining this requires setting the system to prioritize 'novelty' or 'divergence' scores during the retrieval process, effectively force-injecting contradictory evidence into the context window to disrupt confirmation bias. Developers should track the cosine similarity of the embedding vector between consecutive hypothesis iterations; a sustained high-similarity score after a failed experiment is a clear metric for loop stagnation. By focusing on the trajectory of the research state over time, teams can identify when a model is stuck in a loop. Current limitations include high compute overhead for cross-domain vector alignment and the lack of automated hyperparameter tuning for retrieval sensitivity.