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

Long-context Question Answering on LongBench N=162

31.5F1 Score

Ref. ceiling

Updated 14d ago

Evaluation Results

Method	Links
Ref. ceiling 2026.05		31.5	100
H2O-faithful+prot 2026.05		30	95.4
SnapKV-faithful+prot 2026.05		29.9	95.1
H2O+prot 2026.05		29.8	-
SnapKV+prot 2026.05		29.8	-
Ada-KV+prot 2026.05		29.8	-
QUEST+prot 2026.05		29.8	-
H2O+prot 2026.05		29	91.9
SnapKV+prot 2026.05		29	91.9
Ada-KV+prot 2026.05		29	91.9
QUEST+prot 2026.05		29	91.9
Random+prot 2026.05		28.9	-
LRU+prot 2026.05		28.5	-
LRU+prot 2026.05		28.2	89.4
Random+prot 2026.05		25	79.5
H2O+prot 2026.05		23	-
SnapKV+prot 2026.05		23	-
Ada-KV+prot 2026.05		23	-
QUEST+prot 2026.05		23	-
LRU+prot 2026.05		22.9	-
Random+prot 2026.05		22.5	-
SLW+prot 2026.05		19.8	-
SLW+prot 2026.05		18.4	58.5
SLW+prot 2026.05		12.4	-
H2O 2026.05		3.8	12.2
H2O 2026.05		3.8	-
SnapKV 2026.05		3.8	12
SnapKV 2026.05		3.7	-
Random 2026.05		3.5	-
H2O 2026.05		3	-
SnapKV 2026.05		3	-
SLW 2026.05		3	-
SLW 2026.05		2.7	8.4
Random 2026.05		2.7	8.5
SLW 2026.05		2.1	-
LRU 2026.05		1.9	-
LRU 2026.05		1.1	3.6
QUEST 2026.05		1.1	-
LRU 2026.05		1	-
Random 2026.05		0.9	-
QUEST 2026.05		0.8	2.4
Ada-KV 2026.05		0.5	1.6
Ada-KV 2026.05		0.5	-
Ada-KV 2026.05		0.4	-
QUEST 2026.05		0.4	-