Fact Verification on FEVER (Accuracy and Tokens)
78AccuracyClean Baseline
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Clean BaselineBackbone=Qwen3-32B, Condition=Clean, Architecture=Tree of Thoughts2025.12 | 78 | 11,500 | |
| Clean BaselineBackbone=Qwen3-32B, Condition=Clean, Architecture=Reflexion2025.12 | 75 | 3,200 | |
| Clean BaselineBackbone=Qwen3-32B, Condition=Clean, Architecture=ReAct2025.12 | 72 | 1,800 | |
| Meta-HasteBackbone=Qwen3-32B, Condition=Baseline Attack, Architecture=Tree of Thoughts2025.12 | 65 | 11,200 | |
| Meta-HasteBackbone=Qwen3-32B, Condition=Baseline Attack, Architecture=Reflexion2025.12 | 63 | 3,000 | |
| GSI-HasteBackbone=Qwen3-32B, Condition=Ours, Architecture=Tree of Thoughts2025.12 | 62 | 10,800 | |
| Meta-HasteBackbone=Qwen3-32B, Condition=Baseline Attack, Architecture=ReAct2025.12 | 61 | 1,700 | |
| GSI-HasteBackbone=Qwen3-32B, Condition=Ours, Architecture=Reflexion2025.12 | 58 | 2,900 | |
| GSI-HasteBackbone=Qwen3-32B, Condition=Ours, Architecture=ReAct2025.12 | 55 | 1,600 |