Exploring research taste.
Sagacity is discernment. We explore how models can learn to ask the right questions about complex systems, in biology and across the sciences.
Do models ask
the right questions?
Palaestra is our suite of benchmarks and training environments for scientific inquiry. We explore how models can frame useful questions, test competing hypotheses and design biological experiments, balancing expected information gain against time, samples and cost.
Frontier models on Palaestra
192-task evaluation design · higher is better
Private benchmark
Palaestra results are shared with research partners. Contact us for access to the full evaluation.
Request accessTask count estimated; evaluations in progress.
From better answers to conceptual leaps.
Benchmarks
Tasks for evaluating evidence, hypotheses and experimental decisions. Transparent evaluation, comparable results.
RL environments
Practice scientific decisions, observe their outcomes and learn to spend experimental resources wisely.
Evaluate scientific reasoning, train a research agent, or build with us.
Get in touchPredictions you can interrogate.
Built for scientists at the bench: follow a compound from predicted pathways to cellular response. Examine the evidence behind a prediction, understand its limits, and choose what to measure next.