List your company

Zero-shot Question Answering on Reasoning Suite (OBQA, PIQA, ARC-e, ARC-c, Hella, Wino)

46.8Accuracy on OBQA (Zero-shot)Baseline
41.80843.10444.445.696Jun 7, 2026

Evaluation results

MethodLinks
BaselineModel=Mistral-7B-Inst., Compression (%)=02026.0646.880.4181.3155.6383.4874.3570.33
BaselineModel=LLaMA-2-13B, Compression (%)=02026.0644.279.2277.5750.5179.7171.0367.04
ReCalKVModel=Mistral-7B-Inst., Compression (%)=602026.064479.2777.7852.277.8872.367.24
STAR-KVModel=Mistral-7B-Inst., Compression (%)=602026.064480.5279.852.1378.7971.1167.73
STAR-KVModel=LLaMA-2-13B, Compression (%)=602026.0643.67977.3648.1279.371.2766.44
PaluModel=Mistral-7B-Inst., Compression (%)=602026.0642.877.2674.075075.370.7265.03
STAR-KVModel=LLaMA-2-13B, Compression (%)=752026.0642.477.9775.846.1677.7968.3564.75
PaluModel=LLaMA-2-13B, Compression (%)=602026.064276.2871.841.7272.9969.9362.45