Paper Link: https://arxiv.org/pdf/2510.25889

This paper was NOT fun to read, I swear I was getting short breath while trying to understand what the paper was saying

Ultimate skill issue, but I understand Reinforcement Learning and Flow Matching better now

Reliance on SFT introduces critical challenges: curating large-scale, high-quality expert trajectories is laborious and costly

Flow-based Vision-Language-Action Model

Mathematical notation:

Stochasticity Injection

This section talks about converting deterministic denoising process to a probabilistic one

Paper talks about 2 approaches, writing about the Flow-SDE approach

Flow-SDE converts the ODE into an equivalent SDE