Keeping Score: Adaptive, Tuning-Free Loss Weighting for Score-Augmented Neural Ratio Estimation
This work proposes an Adaptive Score Augmented loss (LASA) to improve neural likelihood-to-evidence ratio estimation for structured stochastic process models. By adaptively weighting exact score information into a binary cross-entropy loss, LASA achieves better surrogate quality than standard methods at a fraction of the training data cost, alleviating confidence set undercoverage and improving downstream hypothesis testing.