Role of BCFM regularization as an anchor for learning return distributions

Establish whether the bootstrapped conditional flow matching (BCFM) regularization term functions as an efficient anchor for learning the full return distribution in Value Flows, and characterize the conditions under which this anchoring effect holds.

Background

The practical training objective augments the distributional conditional flow matching loss with a bootstrapped CFM regularization (BCFM) due to instability observed when optimizing the DCFM loss alone.

Empirically, adding BCFM substantially improves performance and sample efficiency in ablations, motivating a conjecture about its anchoring role in learning the return distribution.