Common-sense Reasoning on 5 common-sense reasoning tasks Llama-3-8B
87.07Average AccuracyFULL-SFT
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| FULL-SFTBase Model=Llama-3-8B, Placement=N/A2025.12 | 87.07 | 82.94 | 92.68 | 74.98 | 96.71 | 89.4 | 90.26 | 83.21 | 86.11 | |
| AFA-DORABase Model=Llama-3-8B, Placement=A-B-σ, Decay Configuration=30% linear decay2025.12 | 86.34 | 80.97 | 91.62 | 75.56 | 95.9 | 88.4 | 89.11 | 82.59 | 86.58 | |
| AFA-LoRABase Model=Llama-3-8B, Placement=σ-A-B-σ, Decay Configuration=30% linear decay2025.12 | 86.16 | 81.14 | 91.33 | 75.44 | 95.71 | 88.4 | 88.68 | 81.53 | 87.06 | |
| AFA-DORABase Model=Llama-3-8B, Placement=σ-A-B-σ, Decay Configuration=30% linear decay2025.12 | 86.14 | 81.48 | 91.58 | 76.17 | 95.66 | 86.8 | 89 | 81.42 | 86.97 | |
| AFA-DORABase Model=Llama-3-8B, Placement=σ-A-σ-B-σ, Decay Configuration=30% linear decay2025.12 | 86.13 | 81.65 | 91.07 | 76.6 | 95.7 | 87 | 88.57 | 81.26 | 87.21 | |
| AFA-DORABase Model=Llama-3-8B, Placement=σ-A-σ-B, Decay Configuration=30% linear decay2025.12 | 86.08 | 80.2 | 91.07 | 75.99 | 96.05 | 87.6 | 89.33 | 81.16 | 87.21 | |
| AFA-LoRABase Model=Llama-3-8B, Placement=A-B-σ, Decay Configuration=30% linear decay2025.12 | 86.03 | 81.66 | 91.46 | 74.8 | 95.66 | 87.2 | 89.06 | 81.99 | 86.42 | |
| AFA-LoRABase Model=Llama-3-8B, Placement=A-σ-B-σ, Decay Configuration=30% linear decay2025.12 | 85.89 | 80.72 | 91.41 | 75.08 | 95.87 | 86.4 | 89.17 | 81.06 | 87.37 | |
| AFA-LoRABase Model=Llama-3-8B, Placement=A-σ-B, Decay Configuration=30% linear decay2025.12 | 85.75 | 79.78 | 91.25 | 75.6 | 96.03 | 86.8 | 88.52 | 81.37 | 86.66 | |
| AFA-LoRABase Model=Llama-3-8B, Placement=σ-A-B, Decay Configuration=30% linear decay2025.12 | 85.74 | 81.23 | 91.41 | 75.57 | 95.72 | 86 | 88.08 | 81.93 | 85.95 | |
| AFA-LoRABase Model=Llama-3-8B, Placement=σ-A-σ-B, Decay Configuration=30% linear decay2025.12 | 85.72 | 80.38 | 90.91 | 74.95 | 96.11 | 87.4 | 88.52 | 80.04 | 87.45 | |
| AFA-LoRABase Model=Llama-3-8B, Placement=σ-A-σ-B-σ, Decay Configuration=30% linear decay2025.12 | 85.69 | 80.29 | 90.82 | 75.96 | 95.74 | 87.2 | 89.06 | 80.6 | 85.87 | |
| AFA-DORABase Model=Llama-3-8B, Placement=A-σ-B, Decay Configuration=30% linear decay2025.12 | 85.63 | 79.69 | 91.25 | 75.93 | 95.8 | 86.4 | 88.74 | 80.45 | 86.74 | |
| AFA-DORABase Model=Llama-3-8B, Placement=σ-A-B, Decay Configuration=30% linear decay2025.12 | 85.61 | 80.37 | 91.07 | 75.19 | 95.23 | 85.6 | 89.11 | 81.11 | 87.21 | |
| AFA-DORABase Model=Llama-3-8B, Placement=A-σ-B-σ, Decay Configuration=30% linear decay2025.12 | 85.61 | 79.52 | 90.74 | 75.35 | 95.75 | 85.4 | 89.88 | 81.21 | 87.05 | |
| LoRA BaselineBase Model=Llama-3-8B, Placement=N/A2025.12 | 85.57 | 80.03 | 90.99 | 74.56 | 96.03 | 87 | 88.63 | 81.17 | 86.11 | |
| DORA BaselineBase Model=Llama-3-8B, Placement=N/A2025.12 | 85.45 | 80.46 | 90.45 | 75.66 | 95.77 | 85 | 87.92 | 81.06 | 87.29 | |
| ZO-ActBackbone=Llama-3-8B, q=8, Precision=Full-precision2026.07 | 70.8 | 68.7 | 85.9 | 64.7 | 76.4 | 65.6 | 79.6 | 67.3 | 58.4 | |
| FP16Bits=16, Group=-, Evaluation Protocol=Zero-shot2024.07 | 68.58 | — | — | — | — | — | — | — | — | |
| AGZOBackbone=Llama-3-8B2026.07 | 68.2 | 65 | 83.3 | 64.5 | 70.7 | 60.8 | 79.9 | 63.5 | 57.7 | |
| EfficientQATBits=3, Group=128, Evaluation Protocol=Zero-shot2024.07 | 67.35 | — | — | — | — | — | — | — | — | |
| ZO-ActBackbone=Llama-3-8B, Precision=INT42026.07 | 65.7 | 60.5 | 80.3 | 62.2 | 69.9 | 60.2 | 77 | 61.8 | 53.5 | |
| AWQBits=3, Group=128, Evaluation Protocol=Zero-shot2024.07 | 64.82 | — | — | — | — | — | — | — | — | |
| HiZOOBackbone=Llama-3-8B2026.07 | 64.1 | 60.3 | 79.6 | 63.6 | 65.6 | 58.2 | 73.1 | 58.6 | 53.5 | |
| OmniQBits=3, Group=128, Evaluation Protocol=Zero-shot2024.07 | 64.09 | — | — | — | — | — | — | — | — | |
| SubZeroBackbone=Llama-3-8B2026.07 | 63.9 | 60.2 | 78.3 | 62.5 | 68 | 53.6 | 76.7 | 59.2 | 52.6 | |
| ZO-MuonBackbone=Llama-3-8B, q=82026.07 | 63.1 | 61.4 | 78 | 63 | 63.6 | 50.8 | 76.7 | 57.3 | 53.6 | |
| EfficientQATBits=2, Group=64, Evaluation Protocol=Zero-shot2024.07 | 60.76 | — | — | — | — | — | — | — | — | |
| GPTQBits=3, Group=128, Evaluation Protocol=Zero-shot2024.07 | 60.58 | — | — | — | — | — | — | — | — | |
| EfficientQATBits=2, Group=128, Evaluation Protocol=Zero-shot2024.07 | 59.37 | — | — | — | — | — | — | — | — | |
| RTNBits=3, Group=128, Evaluation Protocol=Zero-shot2024.07 | 58.72 | — | — | — | — | — | — | — | — | |
| LOZOBackbone=Llama-3-8B2026.07 | 57.6 | 51.9 | 69.6 | 62.6 | 48.3 | 53 | 70.2 | 54.5 | 50.6 | |
| OmniQBits=2, Group=128, Evaluation Protocol=Zero-shot2024.07 | 52.66 | — | — | — | — | — | — | — | — | |
| DB-LLMBits=2, Group=64, Evaluation Protocol=Zero-shot2024.07 | 51.74 | — | — | — | — | — | — | — | — | |
| Zero-shotBackbone=Llama-3-8B, Evaluation protocol=Zero-shot2026.07 | 21.2 | 10 | 12.5 | 61.9 | 12.3 | 23.4 | 47.4 | 2.2 | 0 |