Question Answering on HotpotQA (F1 and Cost)
40.9F1 ScoreGreedy cost-aware selection
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Greedy cost-aware selectionBackbone=Qwen3-8B, Bgt/k=152026.06 | 40.9 | 12.6 | |
| Maximum Marginal Relevance (MMR) with cost penaltyBackbone=Qwen3-8B, Bgt/k=202026.06 | 40.77 | 10.6 | |
| Knapsack-based selectionBackbone=Qwen3-8B, Bgt/k=102026.06 | 40.33 | 9.2 | |
| Greedy cost-aware selectionBackbone=Qwen3-8B, Bgt/k=102026.06 | 39.64 | 8.9 | |
| Maximum Marginal Relevance (MMR) with cost penaltyBackbone=Qwen3-8B, Bgt/k=252026.06 | 39.59 | 10.9 | |
| RelevanceBackbone=Qwen3-8B, Bgt/k=252026.06 | 38.7 | 17.2 | |
| Redundancy-aware knapsackBackbone=Qwen3-8B, Bgt/k=202026.06 | 38.66 | 0.2 | |
| Maximum Marginal Relevance (MMR) with cost penaltyBackbone=Qwen3-8B, Bgt/k=302026.06 | 38.35 | 10.9 | |
| RelevanceBackbone=Qwen3-8B, Bgt/k=302026.06 | 38.11 | 18.1 | |
| Maximum Marginal Relevance (MMR) with cost penaltyBackbone=Qwen3-8B, Bgt/k=152026.06 | 37.92 | 9.8 | |
| Vanilla LLMBackbone=Qwen3-8B, Bgt/k=-2026.06 | 37.8 | — | |
| Maximum Marginal Relevance (MMR) with cost penaltyBackbone=Qwen3-8B, Bgt/k=∞2026.06 | 37.71 | 10.9 | |
| Top-kBackbone=Qwen3-8B, Bgt/k=252026.06 | 37.7 | — | |
| Top-kBackbone=Qwen3-8B, Bgt/k=102026.06 | 37.19 | — | |
| Maximum Marginal Relevance (MMR) with cost penaltyBackbone=Qwen3-8B, Bgt/k=102026.06 | 37.17 | 8 | |
| Knapsack-based selectionBackbone=Qwen3-8B, Bgt/k=202026.06 | 37.09 | 15.3 | |
| Knapsack-based selectionBackbone=Qwen3-8B, Bgt/k=302026.06 | 36.93 | 18 | |
| Greedy cost-aware selectionBackbone=Qwen3-8B, Bgt/k=202026.06 | 36.54 | 15 | |
| RelevanceBackbone=Qwen3-8B, Bgt/k=202026.06 | 36.44 | 15.5 | |
| Knapsack-based selectionBackbone=Qwen3-8B, Bgt/k=02026.06 | 36.31 | 0 | |
| RelevanceBackbone=Qwen3-8B, Bgt/k=102026.06 | 35.73 | 9.2 | |
| Greedy cost-aware selectionBackbone=Qwen3-8B, Bgt/k=252026.06 | 35.55 | 16.9 | |
| RelevanceBackbone=Qwen3-8B, Bgt/k=∞2026.06 | 35.49 | 19.2 | |
| Top-kBackbone=Qwen3-8B, Bgt/k=302026.06 | 35.43 | — | |
| RelevanceBackbone=Qwen3-8B, Bgt/k=152026.06 | 35.35 | 12.9 | |
| Greedy cost-aware selectionBackbone=Qwen3-8B, Bgt/k=302026.06 | 35.26 | 18 | |
| Knapsack-based selectionBackbone=Qwen3-8B, Bgt/k=∞2026.06 | 35.18 | 19.2 | |
| Knapsack-based selectionBackbone=Qwen3-8B, Bgt/k=152026.06 | 35.02 | 12.6 | |
| Greedy cost-aware selectionBackbone=Qwen3-8B, Bgt/k=02026.06 | 34.98 | 0 | |
| Knapsack-based selectionBackbone=Qwen3-8B, Bgt/k=252026.06 | 34.97 | 17.1 | |
| Maximum Marginal Relevance (MMR) with cost penaltyBackbone=Qwen3-8B, Bgt/k=52026.06 | 34.63 | 4.7 | |
| Greedy cost-aware selectionBackbone=Qwen3-8B, Bgt/k=∞2026.06 | 34.57 | 19.2 | |
| Maximum Marginal Relevance (MMR) with cost penaltyBackbone=Qwen3-8B, Bgt/k=02026.06 | 34.49 | 0 | |
| Top-kBackbone=Qwen3-8B, Bgt/k=152026.06 | 34.45 | — | |
| RelevanceBackbone=Qwen3-8B, Bgt/k=02026.06 | 34.44 | 0 | |
| Redundancy-aware knapsackBackbone=Qwen3-8B, Bgt/k=302026.06 | 34.21 | 0.2 | |
| Redundancy-aware knapsackBackbone=Qwen3-8B, Bgt/k=∞2026.06 | 34.19 | 0.2 | |
| Redundancy-aware knapsackBackbone=Qwen3-8B, Bgt/k=152026.06 | 33.78 | 0.2 | |
| RelevanceBackbone=Qwen3-8B, Bgt/k=52026.06 | 33.73 | 4.8 | |
| Knapsack-based selectionBackbone=Qwen3-8B, Bgt/k=52026.06 | 33.7 | 4.8 | |
| Greedy cost-aware selectionBackbone=Qwen3-8B, Bgt/k=52026.06 | 33.4 | 4.7 | |
| Redundancy-aware knapsackBackbone=Qwen3-8B, Bgt/k=102026.06 | 33.34 | 0.2 | |
| Top-kBackbone=Qwen3-8B, Bgt/k=202026.06 | 33.02 | — | |
| Redundancy-aware knapsackBackbone=Qwen3-8B, Bgt/k=52026.06 | 32.61 | 0.2 | |
| Redundancy-aware knapsackBackbone=Qwen3-8B, Bgt/k=02026.06 | 32.51 | 0 | |
| Top-kBackbone=Qwen3-8B, Bgt/k=52026.06 | 31.9 | — | |
| Redundancy-aware knapsackBackbone=Qwen3-8B, Bgt/k=252026.06 | 30.47 | 0.2 |