Long-context language evaluation on LongBench benchmark leaderboard · SOTA2 Research

Long-context language evaluation on LongBench

48.43SQABaseline

Evaluation Results

Method	Links
Baseline2025.12		48.43	43.18	25.99	63.22	34.83	59.89	45.92
OmniKV2025.12		48.22	43.05	25.97	63.22	34.72	59.33	45.75
LessIsMore2025.12		48.15	42.71	25.38	63.05	34.67	59.16	45.52
Kascade (All Heads Pooled)2025.12		47.83	40.5	25.34	63.09	34.5	62.95	45.7
Baseline2025.12		47.56	41.35	24.15	64.32	34.83	65.9	46.35
Kascade2025.12		47.41	39.84	25.21	61.32	33.67	62.7	45.02
PBS-Attn2025.10		47.04	37.17	16.66	33.88	66.33	2.8	33.98
Quest2025.12		46.97	42.82	25.71	62.33	34.14	54.36	44.39
Full2025.10		46.93	37.97	16.57	34.07	66	2.93	34.08
MInference2025.10		46.92	37.11	16.53	34.15	66.17	2.98	33.98
FlexPrefill2025.10		46.08	37.29	16.53	33.63	49	2.52	30.84
MeanPooling2025.10		45.69	37.25	16.61	37.45	62.33	2.52	33.64
XAttention2025.10		45.66	37.77	16.66	36.2	64.67	2.02	33.83
Kascade (All Heads Pooled)2025.12		44.87	42.34	23.74	61.99	34.5	62.71	45.02
Quest2025.12		44.71	40.46	24.34	62.72	33.94	57.53	43.95
Kascade2025.12		44.19	40.38	23.02	60.83	35	63.98	44.57
LessIsMore2025.12		40.87	38.47	23.03	62.38	34.67	63.21	43.77
StreamingLLM2025.12		24.83	25.05	22.41	56.33	12	59.89	33.42
StreamingLLM2025.12		24.33	28.44	21.01	56.82	12.5	64.32	34.57