New Preprint Introduces kCGM to Calibrate Generative Models to Target Distributions

kCGM extends our previous work https://arxiv.org/abs/2510.10020, CGM, which finetunes to match the mean rather than the distribution of the target features. Code is available at https://github.com/smithhenryd/cgm/tree/main.

We also test kCGM on protein structure generation and regulatory DNA generation, where it significantly improves the diversity of generated structures and the regulatory activity of generated DNA as predicted by AlphaGenome.

For example, we align a small molecule model to a set of <200 known antibiotics. Directly fine-tuning overfits and increases invalid generations, even with regularization! But kCGM with molecular fingerprints provides antibiotic-like generations while decreasing invalid samples.

More generally in biomolecular generative models, samples can look plausible individually, but get the features you care about wrong in distribution. kCGM treats those features as the calibration target using unbiased gradients of a kernel MMD that works for black-box features.