Samples of the long-horizon environments we build around the empirical research and development of models, molecules and materials. We're starting here, with more coming soon.
Our evaluations are rigorous, realistic and long-horizon, with a research lab in the loop. Built around open-ended research problems, they test whether models can form hypotheses, design and run experiments, interpret results and make progress within resource constraints.
We build our training environments research-first, looking for creative, scalable ways to generate high-quality, novel learning signal that is free of reward hacks. Our environments are built to target the failure modes our evaluations expose.
Research on architectures, training recipes, data, adaptation and efficiency, where an agent forms a hypothesis about a model, runs the experiment that tests it and reads the result.
Powered by Lab-0, our machine-learning lab. Evaluation samples are on FrontierRSI ↗, and training environment samples are available on request.
Research on molecular and protein design, property prediction, docking and synthesis planning, where an agent proposes candidates, predicts how they will behave and decides which to make and test next.
Powered by Lab-1, our medicinal chemistry and protein engineering lab. Evaluation and training environment samples are available on request.
Research on structure prediction, property models and simulation-guided discovery, where an agent proposes structures, predicts their properties and decides which to synthesize and characterize next.
Powered by Lab-2, our solid-state materials and characterization lab. Evaluation and training environment samples are available on request.