Report
Sampler-only change reportedly takes MNIST-Sudoku Hard from 0.8% to 65.7% valid boards solved
A user says they changed only the sampler on SRM's MNIST-Sudoku Hard checkpoint, without retraining.
TLDR
A user reports that changing only the sampler on SRM's MNIST-Sudoku Hard checkpoint raised valid boards solved from 0.8% to 65.7%, without retraining. They suggest continuous diffusion may be better at discrete reasoning than standard sampling indicates.
Combined views
12.2K
3 Sources, first seen ago
