CTEM: A Unified Framework for Density Estimation Across Data Types

A new framework for density estimation called Constant-Target Energy Matching (CTEM) has been developed by researchers. This energy-based approach effectively handles continuous, discrete, and mixed-variable domains. Unlike conventional methods that separately address these data types—where continuous models depend on log-density gradients and discrete models utilize concrete scores that can become unstable at low-probability states—CTEM employs a bounded energy-difference transform instead of typical density-ratio regression. It establishes a sample-only training objective with a constant target of 1. This method allows the learned scalar potential to recover log p without needing partition-function estimation or unbounded ratio regression. CTEM significantly enhances performance across various benchmarks. The research can be found on arXiv under the identifier 2605.09085.