Search papers, labs, and topics across Lattice.
3
0
2
0
A single attention module dynamically switches between linear prefilling and full decoding, achieving $21.9\times$ faster prefilling and Transformer-level long-context recall without the quadratic bottleneck.
Billion-parameter LLMs no longer require global backpropagation to pretrain competitively: sharing a single delayed readout across decoupled modules matches standard zero-shot accuracy while yielding up to 1.44x pipeline throughput.
Modular TTT reveals that the right combination of learning rate and weight decay can significantly boost performance, while deeper architectures may actually degrade results.
