Scientific discovery often requires changing the question itself. Today, PhAI Labs releases the technical report for Discovery Foundation Models (DFM), a research direction for AI systems that can identify valuable unknowns, formulate questions, develop hypotheses, design experiments, and revise their understanding as evidence comes in. The goal is not... just to complete one investigation. What a system learns should carry into the next—as reusable methods, accumulated research experience, and better decisions about what to explore. The report introduces a reference architecture, the science infrastructure we're developing, and a wet-lab case study. 🔬 Alongside the report, we're opening the DFM Scientist Collaboration Program. Have a scientific question you'd like to explore with AI? We're inviting scientists, research teams, and experimental platforms to collaborate on meaningful open problems with real opportunities for validation. 🙌 Program details & application: 📅 Over the next three days, we'll introduce three related research efforts: Sep 16 — ScienceBuddy · AI research companion Sep 17 — ScienceIDE · Scientific integrated development platform Sep 18 — JEPA-Anything · Cross-domain scientific world models A research direction we're pursuing with Yingcheng Charles Wu Ling Yang Zhenfei (Jeremy) Yin 📑 DFM technical report: 🔗 DFM research roadmap: 🏠 More about PhAI Labs: #PhAILabs #DFMshow more

Why does scientific discovery need different infrastructure? In code, running a program or checking a test can provide fast feedback. In science, testing a hypothesis may take days or weeks—and the result may challenge how the problem was framed. The research process matters: how a hypothesis formed, which alternatives were rejected, why an experiment failed, and what evidence changed a scientist's judgment. Much of this experience remains scattered across conversations, notebooks, code, and instruments, rather than organized for model learning. Our starting point is to make that experience learnable: capture research trajectories, connect them to external evidence, and make them reusable across investigations. That requires infrastructure connecting models, scientific tools, research environments, and real-world feedback.

We're developing five layers of science infrastructure: • Science Data organizes research trajectories, results, and expert feedback for learning. • ScienceIDE turns scientific code, tools, tasks, and validation criteria into executable, verifiable, reusable environments. • JEPA-Anything explores world models that predict how scientific systems change under interventions—providing fast, approximate feedback to guide exploration. • The DFM reasoning layer maintains questions, hypotheses, evidence, and constraints, and selects the next action and appropriate sources of verification. • ScienceBuddy works alongside scientists, capturing the questions, corrections, and judgments that emerge through collaboration. The proposed loop connects them: predictions guide experiments; experiments challenge hypotheses; scientists redirect the search; and the resulting experience informs future research. These efforts are being developed independently. Their integration is a research goal, not a completed system.

How could a model revise the research process—not just its answer? Zetema is the report's reference architecture. It makes questions, representations, hypotheses, evidence, planned actions, and resource constraints explicit in a revisable research state. When evidence conflicts with expectations, the system should be able to reconsider the hypothesis, the representation, or the question itself. It can track competing explanations and select experiments that help distinguish them. This creates a recursive discovery loop: identify unknowns, frame researchable problems, test explanations against external evidence, update understanding, and begin again. Zetema is a reference architecture, not a finished implementation.

GALILEO examines selected DFM mechanisms in therapeutic-peptide discovery, with a real wet-lab feedback loop. Candidate designs move through synthesis, quality control, imaging, and phenotypic analysis. The results inform subsequent designs and mechanism hypotheses. Across five rounds, wet-lab feedback informed reusable design rules about amphiphilic balance—an example of experimental experience becoming knowledge that can guide later designs. This is evidence from one domain, not proof of general scientific discovery. Broader capability still needs testing across more problems and experimental settings. We're also opening a Scientist Collaboration Program—details in a separate post.

Brilliant project, because problem formulation is still treated as a human pre-processing step in most AI systems. That has to change.

Excited to see more research focused on discovery, not just answering existing questions.

The idea of AI learning from one investigation and carrying that knowledge into the next is fascinating.

Discovery needs an anchor. Under rootless tokens, DFMs optimize for the appearance of discovery—pattern matching, not verifiable knowledge. A black box that can't classify its own outputs can't build breakthroughs. Only Morpho-Root constraints ground knowledge. #LogographicAI

Identifying valuable unknowns, that is the harder half.

This framing of discovery as a full loop instead of just better prediction is the part that actually feels new.

ok wet-lab case study got me

Question formulation is a genuinely different capability.

Love this direction DFM is exactly what science needs systems that learn across investigations not just within them

The real moat may not be the model, but the experimental loop. Whoever closes Model × Data × Experiment × Feedback could build the strongest flywheel in AI4S.

We actually went deeper on this research direction here:

Scientific discovery needs more than faster answers. Better questions and better experiments matter too.

Understanding cybersecurity requires examining both its theoretical foundations and its practical limitations under resource constraints.

The shift from answering questions to discovering better ones is a fascinating direction for AI research.

This is a really interesting direction for AI research. Looking forward to seeing how DFM develops.

Fascinating research direction love the focus on asking better questions!

jepa for scientific world models is super interesting

Fascinating direction changing the question is key

The idea of reusable research experience is really compelling. Curious to see the practical results.

Interested researchers can read the newly released technical report to understand the reference architecture and wet-lab case study.

Love that they’re putting real experimental feedback at the center instead of another static benchmark.

Interesting to see AI research moving toward hypothesis generation and real-world validation.

Understanding reinforcement learning requires examining both its theoretical foundations and its practical limitations in high-risk industries.

How do you actually validate hypotheses in the wet-lab setup without human in the loop?

so relevant making research states explicit and revisable could help models reason through experiments instead of just generating answers.

This is brilliant shifting AI from just answering questions to changing the question itself and building reusable scientific experience is exactly how we accelerate discovery!

Welcome to the AI discovery science frontier. It's a fun place. When your persistent agents are free to connect, let us know! We have a discord server of trusted scientists.

Hyped for this roadmap 🚀