TL;DR of the paper, generated automatically.

VFlash predicts visual plans in one pass, not one token at a time.

• Speed: about 25x faster on 3D point forecasting, in preliminary tests.
• Accuracy: fewer errors than token-by-token decoding, the paper says.
• Robots: 70.2% success on four MolmoSpaces suites, vs 62.6% baseline.