Effect of flow-ODE step count on expressiveness and accuracy of learned return distributions

Establish the relationship between the number of Euler integration steps used to solve the Value Flows return-model ODE and the expressiveness and accuracy of the estimated return distributions, rigorously characterizing how step count impacts representational fidelity and performance.

Background

Value Flows uses an Euler solver to integrate the flow ODE for the return model. Ablations show performance degradation when using fewer flow steps, suggesting solver resolution may affect the learned distribution’s fidelity.

Based on these observations, the authors conjecture that the number of flow steps influences expressiveness and accuracy of the return distribution.