List your company

Long-context evaluation on RULER 128k (matched-train holdout)

36.1AccuracyStill
16.3421.4726.631.73Jun 5, 2026

Evaluation results

MethodLinks
StillK (Cache budget)=2048, Number of seeds=32026.0636.1
StillK (Cache budget)=1024, Number of seeds=32026.0632.9
StillK (Cache budget)=164, Number of seeds=32026.0631.3
StillK (Cache budget)=328, Number of seeds=32026.0630.2
StillK (Cache budget)=512, Number of seeds=32026.0627.4
KV-DistillK (Cache budget)=1024, Number of seeds=32026.0620.3
KV-DistillK (Cache budget)=2048, Number of seeds=32026.0619.4
KV-DistillK (Cache budget)=512, Number of seeds=32026.0619.4
StillK (Cache budget)=82, Number of seeds=32026.0619.3
KV-DistillK (Cache budget)=328, Number of seeds=32026.0617.9
KV-DistillK (Cache budget)=82, Number of seeds=32026.0617.6
KV-DistillK (Cache budget)=164, Number of seeds=32026.0617.1