Long-form Retrieval-Augmented Generation on LongFact
247.3Information Density (Sci.)ArbGraph
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| ArbGraphBackbone=Qwen3-4B2026.04 | 247.3 | 235.5 | 140.6 | 241.2 | 156.4 | 132.5 | 124.5 | 190.2 | 187.6 | |
| FIREBackbone=Qwen3-4B2026.04 | 229.9 | 232.5 | 107.6 | 240.9 | 155.1 | 107.8 | 93.7 | 219.7 | 177.2 | |
| ArbGraphBackbone=LLaMA-3.1-8B2026.04 | 178.6 | 191 | 176.2 | 195.7 | 174.1 | 177.5 | 171.2 | 176 | 180 | |
| FIREBackbone=LLaMA-3.1-8B2026.04 | 162.4 | 170.1 | 158.9 | 175.2 | 160.2 | 161.6 | 155.1 | 158.1 | 165.7 | |
| CoTBackbone=Qwen3-4B2026.04 | 100.1 | 115.9 | 111.7 | 91.3 | 93.9 | 95.5 | 96.6 | 124 | 104.3 | |
| MCTS-RAGBackbone=Qwen3-4B2026.04 | 97.8 | 120 | 95.6 | 76.6 | 92.8 | 98 | 93.8 | 85.8 | 95.2 | |
| Standard RAGBackbone=Qwen3-4B2026.04 | 79.1 | 99.2 | 83.6 | 77.3 | 77.6 | 72.5 | 73.2 | 86.2 | 81.4 | |
| CRITICBackbone=LLaMA-3.1-8B2026.04 | 79 | 68.7 | 85.7 | 63.9 | 77.6 | 52.5 | 64.2 | 79.1 | 72 | |
| ReActBackbone=Qwen3-4B2026.04 | 72.1 | 70.3 | 81.1 | 62 | 70.5 | 56.9 | 60.2 | 86.4 | 70.6 | |
| CRITICBackbone=Qwen3-4B2026.04 | 71.1 | 78.5 | 81 | 70.6 | 71.4 | 76.1 | 71.9 | 83.8 | 75.5 | |
| ReActBackbone=LLaMA-3.1-8B2026.04 | 70.4 | 66.6 | 76.2 | 61.8 | 72 | 48.1 | 63 | 73.5 | 67.1 | |
| MCTS-RAGBackbone=LLaMA-3.1-8B2026.04 | 61.3 | 61.9 | 62 | 59.5 | 59.5 | 58.3 | 54.6 | 62.9 | 60 | |
| CoTBackbone=LLaMA-3.1-8B2026.04 | 56.9 | 61 | 77.7 | 57.8 | 54.8 | 51.7 | 48.7 | 53.5 | 57.8 | |
| Standard RAGBackbone=LLaMA-3.1-8B2026.04 | 42.7 | 51.3 | 54.3 | 47.6 | 43.4 | 38 | 36.3 | 58.5 | 46.9 |