Search papers, labs, and topics across Lattice.
3
0
6
Optimal learning rate and batch size in MoEs depend directly on the expert activation ratio rather than active or total parameter counts, unlocking predictable hyperparameter transfer even at 1/64 sparsity.
AI-generated literature reviews miss the mark, with less than 6% overlap with expert selections, but the 2026 model shows promise for error-free performance.
AI-generated project proposals can match human quality, but AI reviewers show a bias towards favoring their own outputs.
