Validate applicability of FM inference-time scaling to dataset-based initial distributions (cell trajectories)

Determine whether the Noise Search inference-time compute scaling method for Flow Matching that preserves the linear interpolant can be successfully applied in settings where the initial distribution at t=0, p0, is itself a dataset (for example, cell trajectory data), and empirically validate its performance under such non-Gaussian initial conditions.

Background

The paper introduces inference-time compute scaling methods for Flow Matching that preserve the linear interpolant, including a two-stage RS+NS approach. A key property highlighted is the method’s invariance to the initial noise condition, suggesting potential applicability beyond Gaussian priors commonly used in diffusion models.

Building on Flow Matching’s generality regarding the reference distribution, the authors note that some scientific problems (e.g., modeling cell trajectories) involve initial distributions that are not Gaussian but derived directly from datasets. They explicitly state that validating their method in such settings remains future work, identifying an unresolved empirical question regarding generalization to dataset-based initial conditions.