Model # Params
(M)
(Rec / Conv)
Mem. (M elts)
Recall tasks NIAH-1 NIAH-2 NIAH-3
SWDE FDA SQuAD 1K2K4K8K16K32K 1K2K4K8K16K32K 1K2K4K8K16K32K
Transformer
w. RoPE 340 \(\boldsymbol{\infty}\) / 0.0 42.3 34.5 22.1 1001000.00.00.00.0 1001000.00.00.00.0 71.647.60.00.00.00.0
w. Gate (FoX) 376 \(\boldsymbol{\infty}\) / 0.0 52.5 64.3 30.1 10010032.28.04.20.0 10010010024.011.63.2 95.485.664.211.67.20.0
SSM
GLA 475 12.5 / 0.4 29.0 11.4 30.3 74.625.18.22.20.00.0 91.237.221.43.60.00.0 84.257.120.810.22.30.0
GSA 399 12.5 / 0.0 23.8 14.5 24.9 99.297.190.067.429.611.0 96.698.828.05.11.00.0 60.030.113.51.00.00.0
GDN 475 12.5 / 0.4 29.5 8.3 31.3 99.210099.892.041.822.1 99.292.043.617.86.24.0 92.680.637.85.26.82.5
Mamba-2 382 12.5 / 0.4 25.7 14.9 31.9 99.295.652.212.85.42.8 99.898.068.215.44.43.8 53.453.617.41.82.23.2
SWA 374 12.5 / 0.0 10.0 14.4 29.7 29.811.06.23.41.20.0 36.214.410.23.83.20.0 26.29.27.41.41.80.0
Raven 424 12.5 / 0.0 34.1 22.7 35.4 99.810099.899.899.491.4 98.898.098.881.623.08.8 76.843.613.41.00.00.0