Share your thoughts, 1 month free Claude Pro on usSee more

In-Context Retrieval on RULER (S-NIAH/MK-NIAH Metrics)

100S-NIAH-1 (1K)

Mamba-2

Updated 29d ago

Evaluation Results

Method	Links
Mamba-2 2026.06		100	100	97	55.8	99.6	99.6	62.6	21	59.2	38.6	14.4	29	21.2	21.4
KDA 2026.06		100	100	99.2	70.6	100	100	89	30.6	77.4	63.2	26.2	54	44.2	28
GDN-2 2026.06		100	100	100	97.8	100	100	93	39.2	92	89.8	31.8	72.6	51.4	37.8
CARVE 2026.06		100	100	100	98.4	100	100	94.8	42	93.6	91.4	61.8	76.8	55.2	41.4
Transformer (full attn) 2026.06		100	100	51.2	0	100	100	44.2	0	95.8	94.8	37	75.6	66.6	38.2
Mamba-2 + SWA 2026.06		100	100	51.8	25.4	100	99.6	52.4	25.8	97.8	86.8	48	82	58.6	39
Gated DeltaNet + SWA 2026.06		100	100	47.2	22.4	100	99.8	57.3	25.6	94.8	91.2	47.2	91	78.4	44.8
KDA + SWA 2026.06		100	100	51.8	26.2	100	100	56	23	97.2	93.4	51.6	91.4	84	40.4
Mamba-3 SISO + SWA 2026.06		100	100	49.6	26	100	100	58.2	27.8	95	90.4	44	78.8	65.6	33.6
Mamba-3 MIMO + SWA 2026.06		100	100	49	22.8	100	100	53	27.8	99.4	98.4	54.2	82.4	79	46.6
GDN-2 + SWA 2026.06		100	100	55.2	27.4	100	100	57.9	29.2	99.6	99	55.6	93	84.6	48
CARVE + SWA 2026.06		100	100	56.8	28.4	100	100	61.4	30.6	100	100	63.2	95.8	87.3	50.6
Gated DeltaNet 2026.06		99.8	100	100	97.6	100	100	87.2	32	89.8	54.2	60.6	58	37	27.8