Multi-hop Question Answering on MuSiQue (EM, F1, ∆ EM)
30EM3-Agent Baseline
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| 3-Agent BaselineBackbone=Qwen2.5-7B-Instruct, Number of queries=100, Context=20 mixed supporting/distractor paragraphs per query2026.06 | 30 | 39.6 | 16 | |
| SDA (rank)Backbone=Qwen2.5-7B-Instruct, Number of queries=100, Context=20 mixed supporting/distractor paragraphs per query2026.06 | 30 | 38.7 | 16 | |
| PDEBackbone=Qwen2.5-7B-Instruct, Number of queries=100, Context=20 mixed supporting/distractor paragraphs per query2026.06 | 24 | 34.4 | 10 | |
| CoT De-PolarizationBackbone=Qwen2.5-7B-Instruct, Number of queries=100, Context=20 mixed supporting/distractor paragraphs per query2026.06 | 21 | 30.9 | 7 | |
| NF (No-Filter)Backbone=Qwen2.5-7B-Instruct, Number of queries=100, Context=20 mixed supporting/distractor paragraphs per query2026.06 | 14 | 21.3 | — | |
| PDE-RandomBackbone=Qwen2.5-7B-Instruct, Number of queries=100, Context=20 mixed supporting/distractor paragraphs per query2026.06 | 12 | 19.6 | 2 |