Synthetic Sciences

Synthetic Sciences is an AI research lab developing foundation models and infrastructure for scientific research. The lab focuses on the verification problem in empirical sciences, where evaluating long-horizon research involves tradeoffs among speed, cost, and fidelity, making such work difficult to evaluate, train, and improve reliably. Its stated approach pairs new methods for data generation and verification with algorithms designed around these constraints.

The lab's products include OpenScience, Ascent, and FrontierML, with Atlas also listed among its offerings. Current work spans high-fidelity, long-horizon evaluations testing whether models can form hypotheses, design and execute experiments, interpret failures, and progress under uncertainty and resource constraints, alongside inference-time scaffolding providing tools, persistent experimental state, and search budgets. FrontierML is described as the lab's first public benchmark. Additional work covers in-house post-training infrastructure, curated reinforcement-learning environments for long-horizon scientific research, and experiments with novel RL algorithms, including RL over ultra-long horizons and sample-efficient learning when verification and rollouts are expensive.