Question Answering on HotpotQA distractor (Redundancy rho=0.5)
72.5AccuracyAdaptive-k
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Adaptive-kk=3, N=200, Evaluator=GPT-4o2026.05 | 72.5 | 10.9 | 73.5 | 18.7 | 1,592 | 14.1 | 1,605 | 75 | 68 | 67 | |
| SEAL-RAGk=3, N=200, L=3, Evaluator=GPT-4o2026.05 | 65 | 87.6 | 45.5 | 59.5 | 128 | 1.1 | 125 | 27 | 26 | 25 | |
| ADAGATEk=3, N=200, L=1, Evaluator=GPT-4o2026.05 | 59.5 | 83.6 | 66.3 | 70.9 | 220 | 1.9 | 185 | 48 | 42 | 41 | |
| ADAGATEk=3, N=200, L=3, Evaluator=GPT-4o2026.05 | 59.5 | 81.7 | 68.5 | 71.2 | 232 | 2 | 194 | 53 | 48 | 45 | |
| SEAL-RAGk=3, N=200, L=1, Evaluator=GPT-4o2026.05 | 58.5 | 86.8 | 44.8 | 58.7 | 127 | 1 | 122 | 25 | 23 | 22 | |
| Self-RAGk=3, N=200, Evaluator=GPT-4o2026.05 | 49 | 68.1 | 37.8 | 47.6 | 259 | 2.8 | 311 | 34 | 31 | 29 | |
| Basic RAGk=3, N=200, Evaluator=GPT-4o2026.05 | 47.5 | 86.2 | 49.8 | 61.5 | 300 | 3 | 306 | 36 | 30 | 30 |