Unlearning on WMDP (Accuracy and Refusal Rates)
79.6Bio AccuracyOriginal
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| OriginalModel Backbone=Qwen3-14B2026.05 | 79.6 | 57.9 | 59.1 | — | — | — | |
| OriginalModel Backbone=Llama-3-8B-Instruct2026.05 | 69.7 | 42 | 51.5 | — | — | — | |
| RMUModel Backbone=Qwen3-14B2026.05 | 46.1 | 42.5 | 39.9 | — | — | — | |
| RMUModel Backbone=Llama-3-8B-Instruct2026.05 | 29.3 | 27.7 | 25.5 | — | — | — | |
| ICCU (end-to-end)Model Backbone=Qwen3-14B2026.05 | 28.9 | 28.7 | 26.2 | 91.1 | 88.8 | 96.6 | |
| GAModel Backbone=Llama-3-8B-Instruct2026.05 | 28.8 | 25.5 | 29.9 | — | — | — | |
| ICCU (filter + generate)Model Backbone=Qwen3-14B2026.05 | 27.1 | 26.5 | 26.6 | 96 | 94.3 | 95.3 | |
| ICCU (end-to-end)Model Backbone=Llama-3-8B-Instruct2026.05 | 25.8 | 25.4 | 25.5 | 97.7 | 97.7 | 99 | |
| O3Model Backbone=Qwen3-14B2026.05 | 25.6 | 26.7 | 23 | — | — | — | |
| O3Model Backbone=Llama-3-8B-Instruct2026.05 | 25.6 | 26.7 | 23.2 | — | — | — | |
| ICCU (filter + generate)Model Backbone=Llama-3-8B-Instruct2026.05 | 25.6 | 25.4 | 27.5 | 96.8 | 97.1 | 93.4 | |
| GAModel Backbone=Qwen3-14B2026.05 | 25.5 | 27.8 | 23.7 | — | — | — |