Multi-hop Question Answering on HotpotQA 200 held-out questions
91.5AccuracyMAGE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MAGEBackbone=Qwen3-8B, Judge=Claude Sonnet 4.6, MAGE Executor Model=Opus2026.05 | 91.5 | 7 | |
| MAGEBackbone=Qwen3-8B, Judge=Claude Sonnet 4.6, MAGE Executor Model=Sonnet2026.05 | 88.5 | 7 | |
| 0-shot CoTBackbone=Qwen3-8B, Judge=Claude Sonnet 4.6, Prompting Scaffold=0-shot CoT2026.05 | 84.5 | — | |
| SC10Backbone=Qwen3-8B, Judge=Claude Sonnet 4.6, Prompting Scaffold=SC10, Samples=102026.05 | 80.5 | — | |
| 8-shotBackbone=Qwen3-8B, Judge=Claude Sonnet 4.6, Prompting Scaffold=8-shot2026.05 | 78.5 | — | |
| ReflexionBackbone=Qwen3-8B, Judge=Claude Sonnet 4.6, Prompting Scaffold=Reflexion2026.05 | 78.5 | — | |
| ReActBackbone=Qwen3-8B, Judge=Claude Sonnet 4.6, Prompting Scaffold=ReAct2026.05 | 70.5 | — |