rbio1: Training Scientific Reasoning LLMs with Biological World Models as Soft Verifiers

ABSTRACT

A critical bottleneck in training autonomous agents for biology is the scarcity of high-throughput experimental data for reward modeling. We introduce rbio1, a framework that trains scientific reasoning LLMs using 'world models' of biology as soft verifiers instead of ground-truth wet-lab data. By formulating virtual cell model systems that simulate disease-to-healthy state transitions, our agent iteratively refines its hypotheses against the simulated biological environment. This approach allows language-driven agents to conduct large-scale, multi-step inference and deductive reasoning on cellular perturbation tasks without the prohibitive cost of physical experiments.