Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
2
0
2
Structured pruning of Transformer MLPs is no longer a death sentence for Muon-trained models; BMuon achieves compressibility by making channels compete for a global budget.
Forget about trace norms – generalization in chaotic training regimes hinges on the intricate structure of the full Hessian spectrum and its partial determinants.