Grouped Value Attention: Efficient KV Caching via On-Demand Key Reconstruction

Benchmark Model Rank Results
common-sense-reasoning-on-winograndeGVA + DRoPE d_r=16–Accuracy: 53.35
multiple-choice-question-answering-mcqa-on-arc-cGVA + DRoPE d_r=16–Accuracy: 28.32
multiple-choice-question-answering-mcqa-on-arc-eGVA + DRoPE d_r=16–Accuracy: 63.81
sentence-completion-on-hellaswagGVA + DRoPE d_r=16–Accuracy: 42.69
zero-shot-question-answering-on-openbookqaGVA + DRoPE d_r=16–Accuracy: 33.6