Search papers, labs, and topics across Lattice.
New York University
2
0
4
SumCheck outperforms NTT for high-degree polynomials, but NTT can take the lead for low-degree cases depending on memory constraints鈥攖here's no one-size-fits-all solution.
Seemingly interchangeable optimizers can unlock drastically different representation learning, with Muon achieving 2.3x better spectral scaling than AdamW for rare tokens in identical Transformer architectures.
