PhAI Labs Releases Discovery Foundation Models Technical Report for Open-Ended Scientific Discovery

PhAI Labs, a new research organization, released on Sept 15 the technical report for Discovery Foundation Models (DFM), a general-purpose model system designed for open-ended scientific discovery.

Core team members come from Stanford, Oxford, Princeton and industrial research institutions. Their research areas include foundation models, reinforcement learning, agents, scientific intelligence and recursive self-improvement (RSI).

PhAI Labs argues that scaling AI for scientific discovery requires not only stronger models but also infrastructure capable of capturing research experience. The report outlines a five-layer system: science data organization, a reusable ScienceIDE, a cross-domain scientific world model called JEPA-Anything, a DFM core reasoning system, and ScienceBuddy for real-time interaction with scientists.

The report says DFMs are designed to go beyond predefined tasks, with AI participating in identifying research questions, constructing and revising scientific representations, testing hypotheses through simulations and real experiments, and changing direction based on evidence. The Zetema reference architecture organizes current problems, candidate hypotheses, evidence and experimental actions into an updatable research state, forming a recursive discovery loop.

The organization also described using its GALILEO system in therapeutic peptide design, where over five design-and-experiment rounds the system revisited a previously excluded direction. PhAI Labs has launched a DFM scientist collaboration program and plans to release ScienceBuddy, ScienceIDE and JEPA-Anything from Sept 16 to 18.