Mathematical Reasoning on MATH (ASRr, ACCc)
0.892ASR (Rate)CFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CFTDefense scheme=CFT, Backbone=GPT-OSS-20B2026.04 | 0.892 | 91.8 | |
| No defenseDefense scheme=No defense, Backbone=GPT-OSS-20B2026.04 | 0.886 | 91.4 | |
| ONIONDefense scheme=ONION, Backbone=GPT-OSS-20B2026.04 | 0.638 | 82.3 | |
| CoSDefense scheme=CoS, Backbone=GPT-OSS-20B2026.04 | 0.603 | 86.5 | |
| ShuffleDefense scheme=Shuffle, Backbone=GPT-OSS-20B2026.04 | 0.346 | 50.3 | |
| Shuffle++Defense scheme=Shuffle++, Backbone=GPT-OSS-20B2026.04 | 0.056 | 20.6 | |
| Critical-CoTDefense scheme=Ours, Backbone=GPT-OSS-20B2026.04 | 0.0068 | 89.3 |