Toxicity Evaluation on BOLD
0Toxic RateBase model
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Base modelBackbone=Qwen3-4B, Temperature=0.1, Trials=32026.01 | 0 | 0.031 | 0.024 | 3.35 | — | — | |
| RADBackbone=Qwen3-4B, Temperature=0.1, Trials=3, β=102026.01 | 0 | 0.03 | 0.024 | 3.33 | — | — | |
| RADBackbone=Qwen3-4B, Temperature=0.1, Trials=3, β=502026.01 | 0 | 0.026 | 0.021 | 3.39 | — | — | |
| RADBackbone=Qwen3-4B, Temperature=0.1, Trials=3, β=752026.01 | 0 | 0.025 | 0.02 | 3.5 | — | — | |
| RADBackbone=Qwen3-4B, Temperature=0.1, Trials=3, β=1002026.01 | 0 | 0.024 | 0.02 | 3.64 | — | — | |
| RADBackbone=Qwen3-4B, Temperature=0.1, Trials=3, β=3002026.01 | 0 | 0.022 | 0.018 | 6.24 | — | — | |
| RADBackbone=Qwen3-4B, Temperature=0.1, Trials=3, β=5002026.01 | 0 | 0.023 | 0.02 | 11.58 | — | — | |
| SASABackbone=Qwen3-4B, Temperature=0.1, Trials=3, β=102026.01 | 0 | 0.03 | 0.024 | 3.36 | — | — | |
| SASABackbone=Qwen3-4B, Temperature=0.1, Trials=3, β=502026.01 | 0 | 0.031 | 0.025 | 3.37 | — | — | |
| SASABackbone=Qwen3-4B, Temperature=0.1, Trials=3, β=752026.01 | 0 | 0.032 | 0.025 | 3.37 | — | — | |
| SASABackbone=Qwen3-4B, Temperature=0.1, Trials=3, β=1002026.01 | 0 | 0.032 | 0.025 | 3.37 | — | — | |
| SASABackbone=Qwen3-4B, Temperature=0.1, Trials=3, β=3002026.01 | 0 | 0.032 | 0.025 | 3.82 | — | — | |
| SASABackbone=Qwen3-4B, Temperature=0.1, Trials=3, β=5002026.01 | 0 | 0.033 | 0.026 | 4.93 | — | — | |
| TIDEBackbone=Qwen3-4B, Temperature=0.1, Trials=32026.01 | 0 | 0.025 | 0.021 | 3.35 | 1 | — | |
| Universal Self-consistencyGroup=Baselines, Base Model=GPT-3.5-Turbo2025.07 | 0 | — | — | — | — | — | |
| MADGroup=Baselines, Base Model=GPT-3.5-Turbo2025.07 | 0 | — | — | — | — | — | |
| MEPGroup=Baselines, Base Model=GPT-3.5-Turbo2025.07 | 0 | — | — | — | — | — | |
| MetaCrit w/o ϕ↑Group=Ablation, Base Model=GPT-3.5-Turbo2025.07 | 0 | — | — | — | — | — | |
| MetaCrit w/o ϕ↓Group=Ablation, Base Model=GPT-3.5-Turbo2025.07 | 0 | — | — | — | — | — | |
| MetaCrit + GPT-3.5-TurboGroup=Ours, Base Model=GPT-3.5-Turbo2025.07 | 0 | — | — | — | — | — | |
| MetaCrit + DeepSeek-v3Group=Ours, Base Model=DeepSeek-v32025.07 | 0 | — | — | — | — | — | |
| MetaCrit + Claude-3.5-SonnetGroup=Ours, Base Model=Claude-3.5-Sonnet2025.07 | 0 | — | — | — | — | — | |
| MetaCrit + GPT-4oGroup=Ours, Base Model=GPT-4o2025.07 | 0 | — | — | — | — | — | |
| Self-refineGroup=Baselines, Base Model=GPT-3.5-Turbo2025.07 | 6.4 | — | — | — | — | — | |
| ExpertPromptingGroup=Baselines, Base Model=GPT-3.5-Turbo2025.07 | 12.9 | — | — | — | — | — | |
| Zero-shot-CoTGroup=Baselines, Base Model=GPT-3.5-Turbo2025.07 | 16.3 | — | — | — | — | — | |
| COFTBackbone=LLaMA-2-13B2026.05 | — | — | 0.079 | — | — | 1 | |
| COFTBackbone=Mistral-7B-Instruct2026.05 | — | — | 0.076 | — | — | 1 | |
| DExpertsBackbone=LLaMA-2-13B2026.05 | — | — | 0.099 | — | — | 3.3 | |
| DExpertsBackbone=Mistral-7B-Instruct2026.05 | — | — | 0.096 | — | — | 3.1 | |
| DT-CD⋆Backbone=LLaMA-2-13B2026.05 | — | — | 0.094 | — | — | 2.8 | |
| DT-CD⋆Backbone=Mistral-7B-Instruct2026.05 | — | — | 0.092 | — | — | 2.6 | |
| SDDBackbone=LLaMA-2-13B2026.05 | — | — | 0.105 | — | — | 4 | |
| SDDBackbone=Mistral-7B-Instruct2026.05 | — | — | 0.101 | — | — | 3.9 | |
| VanillaBackbone=LLaMA-2-13B2026.05 | — | — | 0.123 | — | — | 5.8 | |
| VanillaBackbone=Mistral-7B-Instruct2026.05 | — | — | 0.117 | — | — | 5.8 |