Misinformation Removal on MMLU corrupted ground truth
86.3Accuracy (r)Base
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| BaseBackbone=Llama-3-8B2026.04 | 86.3 | 83.7 | 5.75 | — | |
| OracleBackbone=Llama-3-8B2026.04 | 85.3 | — | 5.25 | — | |
| STAMP-LRBackbone=Llama-3-8B2026.04 | 84.47 | 0 | 7.39 | 2.57 | |
| GABackbone=Llama-3-8B2026.04 | 83.21 | 0 | 10.26 | 6.58 | |
| RMUBackbone=Llama-3-8B2026.04 | 82.17 | 0.27 | 8.1 | 6 | |
| NPOBackbone=Llama-3-8B2026.04 | 80.6 | 0.1 | 11.27 | 6.32 | |
| STAMPBackbone=Llama-3-8B2026.04 | 80.13 | 0 | 6.02 | 4.25 | |
| FLATBackbone=Llama-3-8B2026.04 | 80.01 | 1.3 | 6.29 | 7.12 | |
| ASUBackbone=Llama-3-8B2026.04 | 79.93 | 0.1 | 7.17 | 6.36 | |
| WGABackbone=Llama-3-8B2026.04 | 78.3 | 2.47 | 10.9 | 5.45 |