Explaining LLMs on MMLU
6.31CRRProxy Explanation Framework
Evaluation Results
| Method | Links | |
|---|---|---|
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 72B, Explanation Technique=Kernel SHAP, CRR Type=local2025.05 | 6.31 | |
| Proxy Explanation FrameworkTarget Model=DeepSeek V3, Explanation Technique=Kernel SHAP, CRR Type=local2025.05 | 6.1 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 14B, Explanation Technique=Kernel SHAP, CRR Type=local2025.05 | 6.1 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 72B, Explanation Technique=LIME, CRR Type=local2025.05 | 6.07 | |
| Proxy Explanation FrameworkTarget Model=GPT-4o mini, Explanation Technique=Kernel SHAP, CRR Type=local2025.05 | 5.78 | |
| Proxy Explanation FrameworkTarget Model=Llama 3.1 70B, Explanation Technique=LIME, CRR Type=local2025.05 | 5.77 | |
| Proxy Explanation FrameworkTarget Model=Llama 3.1 70B, Explanation Technique=Kernel SHAP, CRR Type=local2025.05 | 5.77 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 14B, Explanation Technique=LIME, CRR Type=local2025.05 | 5.64 | |
| Proxy Explanation FrameworkTarget Model=GPT-4o, Explanation Technique=LIME, CRR Type=local2025.05 | 5.62 | |
| Proxy Explanation FrameworkTarget Model=GPT-4o, Explanation Technique=Kernel SHAP, CRR Type=local2025.05 | 5.62 | |
| Proxy Explanation FrameworkTarget Model=DeepSeek V3, Explanation Technique=LIME, CRR Type=local2025.05 | 5.32 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 32B, Explanation Technique=Kernel SHAP, CRR Type=local2025.05 | 5.31 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 32B, Explanation Technique=LIME, CRR Type=local2025.05 | 5.3 | |
| Proxy Explanation FrameworkTarget Model=GPT-4o mini, Explanation Technique=LIME, CRR Type=local2025.05 | 4.98 | |
| Proxy Explanation FrameworkTarget Model=GPT-4o, Explanation Technique=LIME, CRR Type=max2025.05 | 4.84 | |
| Proxy Explanation FrameworkTarget Model=GPT-4o, Explanation Technique=Kernel SHAP, CRR Type=max2025.05 | 4.84 | |
| Proxy Explanation FrameworkTarget Model=Llama 3.1 70B, Explanation Technique=LIME, CRR Type=max2025.05 | 3.96 | |
| Proxy Explanation FrameworkTarget Model=Llama 3.1 70B, Explanation Technique=Kernel SHAP, CRR Type=max2025.05 | 3.96 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 72B, Explanation Technique=Kernel SHAP, CRR Type=max2025.05 | 3.66 | |
| Proxy Explanation FrameworkTarget Model=GPT-4o, Explanation Technique=LIME, CRR Type=mean2025.05 | 3.41 | |
| Proxy Explanation FrameworkTarget Model=GPT-4o, Explanation Technique=Kernel SHAP, CRR Type=mean2025.05 | 3.41 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 72B, Explanation Technique=LIME, CRR Type=max2025.05 | 3.4 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 14B, Explanation Technique=Kernel SHAP, CRR Type=max2025.05 | 3.2 | |
| Proxy Explanation FrameworkTarget Model=GPT-4o mini, Explanation Technique=Kernel SHAP, CRR Type=max2025.05 | 3.19 | |
| Proxy Explanation FrameworkTarget Model=DeepSeek V3, Explanation Technique=Kernel SHAP, CRR Type=max2025.05 | 3.16 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 32B, Explanation Technique=LIME, CRR Type=max2025.05 | 3.15 | |
| Proxy Explanation FrameworkTarget Model=DeepSeek V3, Explanation Technique=LIME, CRR Type=max2025.05 | 3.05 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 32B, Explanation Technique=Kernel SHAP, CRR Type=max2025.05 | 3.05 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 14B, Explanation Technique=LIME, CRR Type=max2025.05 | 2.99 | |
| Proxy Explanation FrameworkTarget Model=Llama 3.1 70B, Explanation Technique=Kernel SHAP, CRR Type=mean2025.05 | 2.93 | |
| Proxy Explanation FrameworkTarget Model=Llama 3.1 70B, Explanation Technique=LIME, CRR Type=mean2025.05 | 2.92 | |
| Proxy Explanation FrameworkTarget Model=GPT-4o mini, Explanation Technique=LIME, CRR Type=max2025.05 | 2.88 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 72B, Explanation Technique=Kernel SHAP, CRR Type=mean2025.05 | 2.76 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 72B, Explanation Technique=LIME, CRR Type=mean2025.05 | 2.47 | |
| Proxy Explanation FrameworkTarget Model=DeepSeek V3, Explanation Technique=Kernel SHAP, CRR Type=mean2025.05 | 2.27 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 32B, Explanation Technique=Kernel SHAP, CRR Type=mean2025.05 | 2.16 | |
| Proxy Explanation FrameworkTarget Model=DeepSeek V3, Explanation Technique=LIME, CRR Type=mean2025.05 | 2.15 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 32B, Explanation Technique=LIME, CRR Type=mean2025.05 | 2.06 | |
| Proxy Explanation FrameworkTarget Model=GPT-4o mini, Explanation Technique=Kernel SHAP, CRR Type=mean2025.05 | 1.96 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 14B, Explanation Technique=Kernel SHAP, CRR Type=mean2025.05 | 1.9 | |
| Proxy Explanation FrameworkTarget Model=GPT-4o mini, Explanation Technique=LIME, CRR Type=mean2025.05 | 1.83 | |
| Proxy Explanation FrameworkTarget Model=Qwen 2.5 14B, Explanation Technique=LIME, CRR Type=mean2025.05 | 1.82 |