Faithfulness Evaluation on Qasper yes/no question answering
0.102AOPC@10Focus-LIME
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Focus-LIMETarget Model=GPT-4o, Proxy Model=Qwen3-30B-A3B2026.02 | 0.102 | 0.356 | 0.408 | 0.287 | |
| Focus-LIME (w/o Proxy)Target Model=GPT-4o, Use Proxy=false2026.02 | 0.083 | 0.357 | 0.413 | 0.291 | |
| Focus-LIMETarget Model=GPT-4o, Proxy Model=Qwen2.5-32B2026.02 | 0.081 | 0.341 | 0.391 | 0.279 | |
| Focus-LIMETarget Model=DeepSeek V3, Proxy Model=Qwen3-30B-A3B2026.02 | 0.074 | 0.265 | 0.375 | 0.287 | |
| Focus-LIMETarget Model=DeepSeek V3, Proxy Model=Qwen2.5-32B2026.02 | 0.068 | 0.258 | 0.369 | 0.279 | |
| Focus-LIME (w/o Proxy)Target Model=DeepSeek V3, Use Proxy=false2026.02 | 0.064 | 0.276 | 0.384 | 0.273 | |
| Original LIMETarget Model=GPT-4o2026.02 | 0 | 0 | 0 | 0 | |
| Proxy ExplanationTarget Model=GPT-4o, Proxy Model=Qwen3-30B-A3B2026.02 | 0 | 0 | 0 | 0 | |
| Original LIMETarget Model=DeepSeek V32026.02 | 0 | 0 | 0 | 0 | |
| Proxy ExplanationTarget Model=DeepSeek V3, Proxy Model=Qwen3-30B-A3B2026.02 | 0 | 0 | 0 | 0 |