Sci-CoE is a two-stage training method that helps one language model learn to both solve science problems and check those solutions with very little labeled data.
Large language models are great at words, but they struggle to predict what will happen after they act in a changing world.