Reasoning and Shortcut Detection on Three synthetic benchmarks Average (test)
92.5AccuracySART
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| SART2026.03 | 92.5 | 87.9 | 85.5 | 64 | |
| Influence Filtering2026.03 | 76 | 47.7 | 68.4 | — | |
| Data Filtering2026.03 | 74.7 | 54.6 | 64.7 | 69 | |
| Group DRORequires group annotations=true2026.03 | 73.5 | 36 | 61.5 | — | |
| IRMRequires multiple environments=true2026.03 | 73.3 | 34.4 | 61.5 | — | |
| V-RExRequires multiple environments=true2026.03 | 72.1 | 34.9 | 60.6 | — | |
| Self-Consistency2026.03 | 69.1 | 23.3 | 55.9 | — | |
| LfF2026.03 | 69 | 19.4 | 56.3 | — | |
| Standard Fine-Tuning (SFT)2026.03 | 68.8 | 19.4 | 55.9 | — | |
| Focal Loss2026.03 | 68.2 | 18.2 | 55.4 | — | |
| FishrRequires multiple environments=true2026.03 | 68.1 | 16.7 | 56.3 | — | |
| JTT2026.03 | 67.9 | 19 | 54.9 | — |