Search papers, labs, and topics across Lattice.
Affiliation:
2
0
3
Exact incremental BPE achieves $O(\log t)$ amortized append time by weighting centroid search intervals, a significant speedup over prior $O(\log^2 t)$ methods.
Linear attention's blind memory overwriting can be eliminated by framing associative retrieval as optimal Kalman filtering, tracking uncertainty to weight incoming tokens without sacrificing hardware-efficient parallel scans.
