Many scientific discovery problems are essentially a decision-making problem faced with specific rewards in open space. Recently, a team led by the Professor of Linguistics at University College London released Large Discovery Models, LDMs, specifically designed to solve the question of how to design the next experimental lens in the field of scientific discovery.
LDM is a recursive base model architecture that combines generative base models with Gaussian process reward models: the base model relies on historical information and context, continually expanding and reshaping the current effective search space in open space; the Gaussian process builds Bayes reward signals in that search space, based on experimental history, to evaluate the value of each candidate program for the next round of experiments.
Large Discovery Models: Empirically-grounded Model-Based Open