Multi-hop Question Answering on InfoQA Synthetic Multi-hop QA (2–4 hop)
98Avg F1InfoQA
Evaluation results
| Method | Links | |
|---|---|---|
| InfoQAContext Length=0.5k, Type=Context Average, Backbone=Qwen3-14B2025.09 | 98 | |
| InfoQAType=Overall Average, Backbone=Qwen3-14B2025.09 | 86 | |
| InfoQA w/o Pruning Past Reasoning Trace (P.)Type=Overall Average, Backbone=Qwen3-14B2025.09 | 78 | |
| Self Consistency (S-C)Type=Overall Average, Backbone=Qwen3-14B2025.09 | 75 | |
| Chain-of-Thought (CoT)Type=Overall Average, Backbone=Qwen3-14B2025.09 | 73 | |
| ReActType=Overall Average, Backbone=Qwen3-14B2025.09 | 66 | |
| Plan-and-Solve (P&S)Type=Overall Average, Backbone=Qwen3-14B2025.09 | 66 | |
| Self-Ask (S-A)Type=Overall Average, Backbone=Qwen3-14B2025.09 | 65 | |
| InfoQA w/o Capacity-Aware Task Decomposition (D.)Type=Overall Average, Backbone=Qwen3-14B2025.09 | 65 | |
| Self-Refine (S-R)Type=Overall Average, Backbone=Qwen3-14B2025.09 | 57 | |
| InfoQAContext Length=10k, Type=Context Average, Backbone=Qwen3-14B2025.09 | 54 | |
| DirectType=Overall Average, Backbone=Qwen3-14B2025.09 | 32 |