List your company

Long-context retrieval and reasoning on RULER (test)

92.72Retrieval (EM)Llama 3.1 8B Instruct
2.676826.053449.4372.8066Nov 5, 2025

Evaluation results

MethodLinks
Llama 3.1 8B InstructBase Model=Llama 3.1 8B Instruct, KV Cache Management Method=Full Attention, KV Cache Length/Budget=Full2025.1192.7268.0836.0759.13
SnapStreamBase Model=Llama 3.1 8B Instruct, KV Cache Management Method=SnapStream, KV Cache Length/Budget=8k2025.1187.3868.4827.7459.68
SAGE-KVBase Model=Llama 3.1 8B Instruct, KV Cache Management Method=SAGE-KV, KV Cache Length/Budget=8k2025.1160.4365.2424.9258.78
StreamingLLMBase Model=Llama 3.1 8B Instruct, KV Cache Management Method=StreamingLLM, KV Cache Length/Budget=8k2025.116.143.6437.4245.82