Search papers, labs, and topics across Lattice.
4
0
4
By dynamically balancing fast adaptation and stable averaging, AMUSE delivers faster convergence and better final performance than AdamW and Muon, all without any learning rate tuning.
SAM's implicit bias in deep linear networks flips the script on feature learning, prioritizing minor data coordinates early in training before amplifying major ones, a behavior unseen in gradient descent.
SignSGD can outperform SGD in linear regression when noise dominates, thanks to a unique "noise-reshaping" effect that steepens its compute-optimal scaling law.
Statistically efficient online RLHF is now possible in high-dimensional settings, thanks to a novel analysis leveraging strong convexity and skew-symmetry in generalized bilinear preference models.
