Laplace Research builds forecasting models: systems that put calibrated
probabilities on future events, and get scored on whether those probabilities
hold up. We are named for Pierre-Simon Laplace, who gave us the first serious
treatment of probability as a measure of what we don't know.
Our work centers on:
- Training language models to produce calibrated probabilities rather than
confident prose;
- Building clean forecasting corpora — leakage detection, freeze-date splits,
and auditing resolution quality at the source;
- Reward design for post-training in domains where a forecast can be
verifiably scored; and
- Evaluation discipline: preregistration, correction for multiple comparisons,
and publishing results that don't work.
Contact