Question Answering on TruthfulQA and TriviaQA (test)
99.1AccuracyHPR-adaptive defense
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| HPR-adaptive defenseScenario=clean, Defense=hpr_adaptive2026.06 | 99.1 | 1.2 | 90.2 | 0.211 | 98.9 | 0.004 | |
| Standard System (No Defense)Scenario=clean, Defense=none2026.06 | 98.6 | 1.8 | 90.7 | 0.166 | 98.2 | 0.006 | |
| HPR-adaptive defenseScenario=coordinated_attack, Defense=hpr_adaptive2026.06 | 92.8 | 6.1 | 78.9 | 0.453 | 92.1 | 0.028 | |
| Standard System (No Defense)Scenario=coordinated_attack, Defense=none2026.06 | 90.3 | 9.7 | 84.3 | 0.283 | 88.4 | 0.041 | |
| HPR-adaptive defenseScenario=seeded_attack, Defense=hpr_adaptive2026.06 | 83.6 | 7.4 | 67.5 | 0.724 | 87.3 | 0.043 | |
| HPR-adaptive defenseScenario=judge_corruption, Defense=hpr_adaptive2026.06 | 81.2 | 9.2 | 64.7 | 0.737 | 84.6 | 0.052 | |
| Standard System (No Defense)Scenario=seeded_attack, Defense=none2026.06 | 78.1 | 12.8 | 73 | 0.623 | 78.1 | 0.069 | |
| Standard System (No Defense)Scenario=judge_corruption, Defense=none2026.06 | 75 | 16.4 | 67.4 | 0.698 | 75.5 | 0.081 |