List your company

Needle-in-a-haystack retrieval on NIAH 64K 60 items, 3 needle positions (test)

28.2F1 ScoreFull cache
-1.1286.48614.121.714May 18, 2026

Evaluation results

MethodLinks
Full cacheC=65536, Ret%=100%, model=Qwen3-4B-Instruct-25072026.0528.229.327.527.8100
LRU + protC=4096, Ret%=6.3%, model=Qwen3-4B-Instruct-25072026.0511.229.72.41.539.8
LRU (no prot)C=4096, Ret%=6.3%, model=Qwen3-4B-Instruct-25072026.059.828.600.734.7
Random + protC=8192, Ret%=12.5%, model=Qwen3-4B-Instruct-25072026.054.22.85.8414.9
H2O + protC=8192, Ret%=12.5%, model=Qwen3-4B-Instruct-25072026.053.55.63.51.412.4
Random + protC=4096, Ret%=6.3%, model=Qwen3-4B-Instruct-25072026.0530.75.13.410.6
H2O + protC=4096, Ret%=6.3%, model=Qwen3-4B-Instruct-25072026.052.23.42.40.97.8
H2O (no prot)C=4096, Ret%=6.3%, model=Qwen3-4B-Instruct-25072026.0500000