List your company

Question Answering and Commonsense Reasoning on ARC-C, ARC-E, PIQA, Winogrande, HellaSwag

63.48ARC-C AccuracyFP16/BF16
22.129632.864843.654.3352Feb 3, 2026Feb 15, 2026Feb 28, 2026Mar 13, 2026Mar 25, 2026Apr 7, 2026Apr 20, 2026

Evaluation results

MethodLinks
FP16/BF16Backbone=Llama-3.1-70B-Instruct, bit/param=162026.0463.4883.9283.9579.0184.5878.99
QuIPBackbone=Llama-3.1-70B-Instruct, bit/param=3.252026.0462.1282.4582.178.0682.8377.51
GSQBackbone=Llama-3.1-70B-Instruct, bit/param=3.132026.0462.1282.8382.7578.9383.377.99
EfficientQATBackbone=Llama-3.1-70B-Instruct, bit/param=3.252026.0461.8683.8882.4876.0182.7977.4
QTIPBackbone=Llama-3.1-70B-Instruct, bit/param=3.002026.0461.7782.7983.7978.384.2178.17
QTIPBackbone=Llama-3.1-70B-Instruct, bit/param=2.002026.0461.6981.6982.4377.5182.9577.25
GPTQBackbone=Llama-3.1-70B-Instruct, bit/param=3.252026.0460.7580.6483.377.9882.3777.01
GSQBackbone=Llama-3.1-70B-Instruct, bit/param=2.37^NU2026.0460.281.482.480.682.677.4
GSQBackbone=Llama-3.1-70B-Instruct, bit/param=2.62^NU2026.0459.681.482.980.482.977.5
GSQBackbone=Llama-3.1-70B-Instruct, bit/param=2.132026.0458.8779.5581.0776.2482.1175.57
PV-TuningBackbone=Llama-3.1-70B-Instruct, bit/param=2.072026.0458.6280.7281.5677.7482.7276.27
FP16/BF16Backbone=Llama-3.1-8B-Instruct, bit/param=162026.0455.1279.6380.8573.879.1673.71
EfficientQATBackbone=Llama-3.1-70B-Instruct, bit/param=2.372026.0454.8677.2780.3665.6779.0171.43
QTIPBackbone=Llama-3.1-8B-Instruct, bit/param=3.002026.0453.9279.4280.2572.1478.372.81
EfficientQATBackbone=Llama-3.1-8B-Instruct, bit/param=3.252026.0452.9978.9179.7671.5976.8572.02
GSQBackbone=Llama-3.1-8B-Instruct, bit/param=3.132026.0452.9978.3780.0373.5676.6672.32
QuIPBackbone=Llama-3.1-8B-Instruct, bit/param=3.252026.0452.376.6878.8471.975.3771.02
QTIPBackbone=Llama-3.1-8B-Instruct, bit/param=2.002026.0450.6875.4278.1870.0975.0269.88
GSQBackbone=Llama-3.1-8B-Instruct, bit/param=2.37^NU2026.0450.2674.4178.1369.6174.8169.44
PV-TuningBackbone=Llama-3.1-8B-Instruct, bit/param=2.272026.0450.2673.9179.1670.5675.2869.83
GSQBackbone=Llama-3.1-8B-Instruct, bit/param=2.132026.0448.1272.3578.0770.873.4268.55
Full CacheBackbone=Llama-3-8B, Rate=100%, Evaluation Protocol=Zero-shot2026.0245.7670.0280.8571.5174.6868.56
DynaKVBackbone=Llama-3-8B, Rate=47%, Evaluation Protocol=Zero-shot2026.0244.7567.3780.3667.6471.766.36
Full CacheBackbone=Qwen2-7B-Base, Rate=100%, Evaluation Protocol=Zero-shot2026.0244.4171.679.7670.472.2867.69
EfficientQATBackbone=Llama-3.1-8B-Instruct, bit/param=2.372026.0443.7767.5574.6564.3368.6563.79
GSQBackbone=Llama-3.1-8B-Instruct, bit/param=1.712026.0442.8367.1373.565.8267.9163.44
DynaKVBackbone=Llama-3-8B, Rate=29%, Evaluation Protocol=Zero-shot2026.0242.7159.9679.8266.5470.3363.87
DynaKVBackbone=Qwen2-7B-Base, Rate=48%, Evaluation Protocol=Zero-shot2026.0242.3770.1978.7864.9668.1864.9
DynaKVBackbone=Qwen2-7B-Base, Rate=30%, Evaluation Protocol=Zero-shot2026.024050.7976.7163.1467.0859.54
GPTQBackbone=Llama-3.1-8B-Instruct, bit/param=3.252026.0439.6858.6767.0366.364.2459.18
QuIPBackbone=Llama-3.1-70B-Instruct, bit/param=2.372026.0439.5160.4473.6165.3568.9561.57
PaluBackbone=Llama-3-8B, Rate=50%, Evaluation Protocol=Zero-shot2026.0239.3263.8477.5867.86963.51
DynaKVBackbone=Qwen2-7B-Base, Rate=18%, Evaluation Protocol=Zero-shot2026.0239.3259.9675.4158.7262.4459.17
DynaKVBackbone=Llama-3-8B, Rate=20%, Evaluation Protocol=Zero-shot2026.0238.9859.4478.5665.5167.8962.08
MatryoshkaKVBackbone=Llama-3-8B, Rate=50%, Evaluation Protocol=Zero-shot2026.0238.6460.4978.0263.1467.3661.53
GPTQBackbone=Llama-3.1-70B-Instruct, bit/param=2.372026.0438.2358.6372.857.1460.1157.38
MatryoshkaKVBackbone=Llama-3-8B, Rate=30%, Evaluation Protocol=Zero-shot2026.0232.8849.7473.6755.856.853.78
PaluBackbone=Qwen2-7B-Base, Rate=50%, Evaluation Protocol=Zero-shot2026.0232.8849.7471.3862.4360.7455.43
PaluBackbone=Llama-3-8B, Rate=20%, Evaluation Protocol=Zero-shot2026.0229.4928.5765.7254.746.4744.99
PaluBackbone=Llama-3-8B, Rate=30%, Evaluation Protocol=Zero-shot2026.0229.1527.5171.660.6257.1949.21
MatryoshkaKVBackbone=Llama-3-8B, Rate=20%, Evaluation Protocol=Zero-shot2026.0227.4645.1568.553.8345.3148.05
PaluBackbone=Qwen2-7B-Base, Rate=30%, Evaluation Protocol=Zero-shot2026.0227.1237.7461.9753.7540.0444.12
PaluBackbone=Qwen2-7B-Base, Rate=20%, Evaluation Protocol=Zero-shot2026.0226.4431.5756.9150.9932.3939.66
GPTQBackbone=Llama-3.1-8B-Instruct, bit/param=2.372026.0424.9127.6152.8351.7830.537.53
QuIPBackbone=Llama-3.1-8B-Instruct, bit/param=2.372026.0423.7228.9653.8650.8338.6539.2