New Theory Explains the Effectiveness of Stop-Gradient in Flow Models

kwangmoo_yi · x · 2026-08-06

The paper "Flow Map Learning via Nongradient Vector Flow" by Goldstein et al. explores the use of stop-gradient in training flow models. The author notes that while placing stop-gradient has previously been shown to be useful in practice, this research provides a theoretical foundation explaining why it works.

Related event: ICLR Paper Introduces SGFlow for Efficient Flow Matching Without Backpropagation(2 posts)→

Original post →

More from Research

Research channel →