List your company

Common-sense QA on PIQA

79AccuracyFP16 (no compression)
72.999274.557176.11577.6729Aug 13, 2023Jan 10, 2024Jun 8, 2024Nov 5, 2024Apr 4, 2025Sep 1, 2025Jan 29, 2026

Evaluation results

MethodLinks
FP16 (no compression)error budget (epsilon)=N/A, KV cache (rel.)=1, Zero-shot protocol=true, Base Model=Llama-3-8B2026.0179—
StiefAttentionerror budget (epsilon)=0.015, KV cache (rel.)=0.85, Zero-shot protocol=true, Base Model=Llama-3-8B2026.0179—
FP16Backbone=LLaMA2-7B, Avg. Bit=16, Evaluation Protocol=zero-shot2024.1078.07—
StiefAttentionerror budget (epsilon)=0.03, KV cache (rel.)=0.76, Zero-shot protocol=true, Base Model=Llama-3-8B2026.0178—
CMPQBackbone=LLaMA2-7B, Avg. Bit=4, Evaluation Protocol=zero-shot2024.1077.93—
QuIPBackbone=LLaMA2-7B, Avg. Bit=4, Evaluation Protocol=zero-shot2024.1077.73—
RTNBackbone=LLaMA2-7B, Avg. Bit=4, Evaluation Protocol=zero-shot2024.1077.58—
GPTQBackbone=LLaMA2-7B, Avg. Bit=4, Evaluation Protocol=zero-shot2024.1077.48—
AWQBackbone=LLaMA2-7B, Avg. Bit=4, Evaluation Protocol=zero-shot2024.1077.31—
OPT-2.7B FP16Model=OPT-2.7B, Precision=FP162023.0876.7110.91
StiefAttentionerror budget (epsilon)=0.045, KV cache (rel.)=0.67, Zero-shot protocol=true, Base Model=Llama-3-8B2026.0176—
CMPQBackbone=LLaMA2-7B, Avg. Bit=3, Evaluation Protocol=zero-shot2024.1075.7—
TSLDModel=OPT-2.7B, Quantization=QAT-KD (tensor-wise)2023.0875.6211.35
QuIPBackbone=LLaMA2-7B, Avg. Bit=3, Evaluation Protocol=zero-shot2024.1075.12—
StiefAttentionerror budget (epsilon)=0.06, KV cache (rel.)=0.61, Zero-shot protocol=true, Base Model=Llama-3-8B2026.0175—
GT+LogitModel=OPT-2.7B, Quantization=QAT-KD (tensor-wise)2023.0874.9712.1
LogitModel=OPT-2.7B, Quantization=QAT-KD (tensor-wise)2023.0874.3211.69
StiefAttentionerror budget (epsilon)=0.09, KV cache (rel.)=0.57, Zero-shot protocol=true, Base Model=Llama-3-8B2026.0174—
RTNBackbone=LLaMA2-7B, Avg. Bit=3, Evaluation Protocol=zero-shot2024.1073.28—
GPTQBackbone=LLaMA2-7B, Avg. Bit=3, Evaluation Protocol=zero-shot2024.1073.23—