Tagged articles

backpropagation-free

1 articles · Page 1 of 1
Machine Heart
Machine Heart
Oct 25, 2026 · Artificial Intelligence

Dust: Zeroth-Order Optimization Trains Transformers Without Backpropagation

Q Labs introduces Dust, a zeroth-order optimization method that trains Transformers by adding Gaussian noise to per-token activations instead of using backpropagation, achieving competitive results at 243M parameters and 20M tokens while challenging the necessity of differentiability in neural network training.

Q LabsTransformer pretrainingactivation perturbation
0 likes · 15 min read
Dust: Zeroth-Order Optimization Trains Transformers Without Backpropagation