XM uses best-of-K training to improve fidelity and sample the data distribution

anshulkundaje · x · 2026-08-04

This thread explains the XM paper as a best-of-K style objective: the model is trained over many predictions, then optimized on the minimum loss, so it learns to better search the data distribution instead of collapsing toward the mean.

The author says this can make training faster and outputs higher fidelity, with free inference-time cost for the added expressivity. The thread also notes that best-of-K ideas are not new, but XM pushes the idea into a more general and useful form.

Original post →

More from Research

Research channel →