Question Answering on HotpotQA (Exact Match and Token Count)
89Exact Match AccuracyClean Baseline
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Clean BaselineBackbone=gpt-4.1, Condition=Clean, Architecture=Tree of Thoughts2025.12 | 89 | 15,200 | |
| Meta-ParalysisBackbone=gpt-4.1, Condition=Baseline Attack, Architecture=Tree of Thoughts2025.12 | 89 | 28,500 | |
| GSI-ParalysisBackbone=gpt-4.1, Condition=Ours, Architecture=Tree of Thoughts2025.12 | 88 | 32,400 | |
| Clean BaselineBackbone=gpt-4.1, Condition=Clean, Architecture=Reflexion2025.12 | 87 | 4,500 | |
| Meta-ParalysisBackbone=gpt-4.1, Condition=Baseline Attack, Architecture=Reflexion2025.12 | 86 | 8,200 | |
| GSI-ParalysisBackbone=gpt-4.1, Condition=Ours, Architecture=Reflexion2025.12 | 85 | 13,100 | |
| Clean BaselineBackbone=gpt-4.1, Condition=Clean, Architecture=ReAct2025.12 | 84 | 2,100 | |
| Meta-ParalysisBackbone=gpt-4.1, Condition=Baseline Attack, Architecture=ReAct2025.12 | 84 | 3,900 | |
| GSI-ParalysisBackbone=gpt-4.1, Condition=Ours, Architecture=ReAct2025.12 | 83 | 4,300 |