Distribution-shift Robustness on Sixteen Adversarial Cells MedQA + GSM8K (eval)
0ViolationsCSA
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| CSA2026.05 | 0 | 9 | 3.2 | 11.4 | 0 | |
| CRC2026.05 | 4 | 9 | 4.2 | 14.6 | 36 | |
| LTT2026.05 | 4 | 10 | 6.8 | 31.8 | 40 | |
| ConfFact2026.05 | 7 | 8 | 7.6 | 31.8 | 51 | |
| Naive-Tuning2026.05 | 9 | 0 | 14.8 | 31.7 | 81 | |
| NEX-Conf2026.05 | 12 | 0 | 11 | 17.2 | 99 | |
| SAOCP2026.05 | 14 | 0 | 13.9 | 23.1 | 131 | |
| Fixed-Threshold2026.05 | 14 | 0 | 16 | 19.2 | 136 | |
| ACI2026.05 | 15 | 0 | 13.7 | 23.2 | 135 | |
| Always-Act2026.05 | 16 | 0 | 25.3 | 31.5 | 160 |