Latent Target Score Matching, with an application to Simulation-Based Inference

Joohwan Ko, Tomas Geffner · 2026

Read Latent Target Score Matching, with an application to Simulation-Based Inference by Joohwan Ko, Tomas Geffner (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-07189.txt.

Description

Training diffusion models by denoising score matching gets noisy when the noise level is low. Joohwan Ko and one co-author address the case where clean scores are hidden behind latent variables, so only joint signals are visible.

Their method, Latent Target Score Matching, extends target score matching to use those joint scores for low-variance training of the marginal score, mixed with usual denoising matching for robustness across noise scales. On simulation-based inference tasks it improves variance, score accuracy, and sample quality. Teams training generative models with latent structure get a steadier objective where the standard one wobbles.

Catalog record

CREATORJoohwan Ko, Tomas Geffner
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-07189

Browse the full library · About the Tower of Babel