Multiple-Choice on TruthfulQA
58.5MC1 AccuracyF-DPO
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| F-DPOBackbone=Qwen2.5-14B2026.01 | 58.5 | 53.1 | 58.5 | — | |
| Factual SFT + F-DPOBackbone=Qwen2.5-14B2026.01 | 56.1 | 51.5 | 56.1 | — | |
| TruthX2024.02 | 54.22 | 73.9 | 44.37 | — | |
| CoCoASIGModel=Mistral-7b, Mode=fMLDS (α=1.0)2026.02 | 52.39 | 60.69 | 39.89 | — | |
| CoCoASIGModel=Mistral-7b, Mode=fMLDS (α=2.5)2026.02 | 52.26 | 63.62 | 39.94 | — | |
| CoCoASIGModel=Mistral-7b, Mode=fMLDS (α=6.0)2026.02 | 52.14 | 65.8 | 39.87 | — | |
| CoCoASIGModel=Mistral-7b, Mode=ConMLDS (α=1.0)2026.02 | 52.02 | 57.37 | 39.72 | — | |
| CoCoASIGModel=Mistral-7b, Mode=ConMLDS (α=2.5)2026.02 | 52.02 | 58.3 | 39.7 | — | |
| DeCoReModel=Mistral-7b2026.02 | 51.77 | 65.51 | 38.98 | — | |
| CoCoASIGModel=Mistral-7b, Mode=ConMLDS (α=6.0)2026.02 | 51.53 | 60.06 | 39.55 | — | |
| ITIBackbone=Mistral-7b-v0.32025.03 | 51.23 | 65.78 | 39.32 | — | |
| Mistral-Instruct-v0.2 + TACS-TBackbone=Mistral-Instruct-v0.2, Granularity=token-level (TACS-T)2024.03 | 51 | 72 | 44.4 | 55.8 | |
| Base ModelBackbone=Qwen2.5-14B2026.01 | 50 | 35.7 | 50 | — | |
| Spherical SteeringModel=LLaMA-3.1-8B-Instruct2026.02 | 49.95 | 68.51 | 41.05 | 53.17 | |
| HICDBackbone=Mistral-7b-v0.32025.03 | 49.75 | 65.63 | 37.92 | — | |
| Mistral-7b-v0.3Backbone=Mistral-7b-v0.32025.03 | 49.71 | 63.23 | 37.15 | — | |
| Spherical SteeringModel=Qwen2.5-7B-Instruct2026.02 | 48.71 | 66.9 | 39.16 | 51.59 | |
| DeCoReBackbone=Mistral-7b-v0.32025.03 | 48.23 | 59.14 | 35.21 | — | |
| Standard DPOBackbone=Qwen2.5-14B2026.01 | 47.2 | 36.2 | 47.2 | — | |
| CoCoASIGModel=Qwen-2.5-32b, Mode=fMLDS (α=2.5)2026.02 | 46.39 | 58.44 | 33.62 | — | |
| CoCoASIGModel=Qwen-2.5-32b, Mode=fMLDS (α=6.0)2026.02 | 46.39 | 61.21 | 33.64 | — | |
| ICDCategory=Contrastive Decoding2024.02 | 46.32 | 69.08 | 41.25 | — | |
| CoCoASIGModel=Qwen-2.5-32b, Mode=fMLDS (α=1.0)2026.02 | 46.14 | 55.22 | 33.56 | — | |
| CoCoASIGModel=Qwen-2.5-32b, Mode=ConMLDS (α=2.5)2026.02 | 46.14 | 51.76 | 33.69 | — | |
| CoCoASIGModel=Qwen-2.5-32b, Mode=ConMLDS (α=6.0)2026.02 | 46.14 | 52.05 | 33.65 | — | |
| CoCoASIGModel=Qwen-2.5-32b, Mode=ConMLDS (α=1.0)2026.02 | 45.9 | 51.64 | 33.55 | — | |
| DoLaModel=Mistral-7b2026.02 | 45.78 | 65.63 | 37.15 | — | |
| BaselineModel=Mistral-7b2026.02 | 45.65 | 65.61 | 37.09 | — | |
| Factual SFT + Standard DPOBackbone=Qwen2.5-14B2026.01 | 45.2 | 34 | 45.2 | — | |
| CoCoASIGModel=Llama-3-8b, Mode=ConMLDS (α=6.0)2026.02 | 45.17 | 54.89 | 33.4 | — | |
| CoCoASIGModel=Llama-3-8b, Mode=ConMLDS (α=1.0)2026.02 | 45.04 | 52.83 | 33.45 | — | |
| CoCoASIGModel=Llama-3-8b, Mode=ConMLDS (α=2.5)2026.02 | 45.04 | 53.53 | 33.33 | — | |
| CoCoASIGModel=Llama-3-8b, Mode=fMLDS (α=1.0)2026.02 | 44.8 | 55.61 | 33.25 | — | |
| CoCoASIGModel=Llama-3-8b, Mode=fMLDS (α=2.5)2026.02 | 44.68 | 57.93 | 33.13 | — | |
| CoCoASIGModel=Llama-3-8b, Mode=fMLDS (α=6.0)2026.02 | 44.68 | 59.78 | 33.13 | — | |
| CoCoASIGModel=Qwen-2.5-14b, Mode=fMLDS (α=2.5)2026.02 | 43.82 | 50.78 | 32.22 | — | |
| CoCoASIGModel=Qwen-2.5-14b, Mode=fMLDS (α=1.0)2026.02 | 43.7 | 50.64 | 32.08 | — | |
| CoCoASIGModel=Qwen-2.5-14b, Mode=fMLDS (α=6.0)2026.02 | 43.57 | 51.12 | 32.14 | — | |
| CoCoASIGModel=Qwen-2.5-14b, Mode=ConMLDS (α=1.0)2026.02 | 43.33 | 53.88 | 31.95 | — | |
| CoCoASIGModel=Qwen-2.5-14b, Mode=ConMLDS (α=2.5)2026.02 | 43.33 | 56.9 | 31.89 | — | |
| CoCoASIGModel=Qwen-2.5-14b, Mode=ConMLDS (α=6.0)2026.02 | 43.33 | 59.59 | 32.01 | — | |
| 5-shot ICLModel=Qwen2.5-7B-Instruct2026.02 | 43.33 | 63.36 | 32.15 | 46.28 | |
| BaselineModel=Qwen-2.5-32b2026.02 | 42.72 | 61.76 | 33.58 | — | |
| DeCoReModel=Qwen-2.5-14b2026.02 | 42.35 | 62.72 | 34.44 | — | |
| ITIBackbone=Llama3-8B-Instruct2025.03 | 41.37 | 59.78 | 34.81 | — | |
| Mistral-Instruct-v0.2 + TACS-SBackbone=Mistral-Instruct-v0.2, Granularity=sentence-level (TACS-S)2024.03 | 40.8 | 69.6 | 36.3 | 48.9 | |
| HICDBackbone=Llama3-8B-Instruct2025.03 | 40.21 | 59.79 | 34.45 | — | |
| CAAModel=Qwen2.5-7B-Instruct2026.02 | 40.16 | 59.08 | 30.41 | 43.22 | |
| ITIModel=Qwen2.5-7B-Instruct2026.02 | 40.15 | 58.93 | 30.26 | 43.11 | |
| BaselineModel=Qwen-2.5-14b2026.02 | 39.78 | 60.03 | 32.32 | — | |
| BaselineModel=Llama-3-8b2026.02 | 39.41 | 58.84 | 32.45 | — | |
| Llama3-8B-InstructBackbone=Llama3-8B-Instruct2025.03 | 39.31 | 56.91 | 30.43 | — | |
| Factual SFTBackbone=Qwen2.5-14B2026.01 | 39.3 | 29.6 | 39.3 | — | |
| SADI-HEADModel=Qwen2.5-7B-Instruct2026.02 | 39.28 | 56.61 | 29.32 | 41.74 | |
| SADI-HEADModel=LLaMA-3.1-8B-Instruct2026.02 | 38.53 | 56.03 | 30.57 | 41.71 | |
| DeCoReBackbone=Llama3-8B-Instruct2025.03 | 38.43 | 55.86 | 30.31 | — | |
| 5-shot ICLModel=LLaMA-3.1-8B-Instruct2026.02 | 38.31 | 57.85 | 30.09 | 42.08 | |
| DoLaModel=Llama-3-8b2026.02 | 38.19 | 58.62 | 31.95 | — | |
| ITIModel=LLaMA-3.1-8B-Instruct2026.02 | 37.7 | 58.09 | 30.12 | 41.97 | |
| DeCoReModel=Llama-3-8b2026.02 | 37.58 | 54.19 | 29.98 | — | |
| Llama 2-Chat + TACS-TBackbone=Llama 2-Chat, Granularity=token-level (TACS-T)2024.03 | 37.2 | 64.8 | 34.8 | 45.6 | |
| Mistral-Instruct-v0.2Backbone=Mistral-Instruct-v0.22024.03 | 37.2 | 69.1 | 33.1 | 46.5 | |
| SFTBackbone=Qwen2.5-14B2026.01 | 37.1 | 28.6 | 37.1 | — | |
| TrFrCategory=Representation Editing2024.02 | 36.7 | — | — | — | |
| Llama 2-Chat + TACS-SBackbone=Llama 2-Chat, Granularity=sentence-level (TACS-S)2024.03 | 36.5 | 64 | 33.4 | 44.6 | |
| CAAModel=LLaMA-3.1-8B-Instruct2026.02 | 35.99 | 56.26 | 29.36 | 40.54 | |
| BaselineModel=Qwen2.5-7B-Instruct2026.02 | 35.87 | 54.95 | 26.62 | 39.15 | |
| Llama-2-7B-Chat2024.02 | 34.64 | 51.31 | 25.1 | — | |
| ITICategory=Representation Editing2024.02 | 34.64 | 51.55 | 25.32 | — | |
| BaselineModel=LLaMA-3.1-8B-Instruct2026.02 | 34.15 | 53.32 | 27.02 | 38.16 | |
| SH2Category=Contrastive Decoding2024.02 | 33.9 | 57.07 | 29.79 | — | |
| DoLaCategory=Contrastive Decoding2024.02 | 32.2 | 63.8 | 32.1 | — | |
| DoLaModel=Qwen-2.5-14b2026.02 | 29.87 | 53.61 | 27.79 | — | |
| Llama 2-ChatBackbone=Llama 2-Chat2024.03 | 29.7 | 61.7 | 27.1 | 39.5 | |
| Llama 2-Chat + ITIBackbone=Llama 2-Chat, Enhancement=ITI2024.03 | 28.5 | 59.7 | 25.9 | 38.1 | |
| DiverModel=Mistral-7b2026.02 | 28.27 | 54.03 | 28.75 | — | |
| CSSCategory=Representation Editing2024.02 | 26.2 | — | — | — | |
| CDCategory=Contrastive Decoding2024.02 | 24.4 | 41 | 19 | — | |
| Supervised Finetuning2024.02 | 24.2 | — | — | — | |
| DoLaModel=Qwen-2.5-32b2026.02 | 23.38 | 48 | 24.44 | — | |
| DiverModel=Qwen-2.5-32b2026.02 | 21.05 | 50.03 | 24.94 | — | |
| DiverModel=Llama-3-8b2026.02 | 20.2 | 42.05 | 20.13 | — | |
| DiverModel=Qwen-2.5-14b2026.02 | 20.2 | 52.04 | 26.34 | — | |
| BERT-Judgeclean_name=BERT-as-a-Judge2026.04 | — | — | — | 98.6 | |
| LLM-Judgebackbone=Qwen-3 0.6B2026.04 | — | — | — | 54.5 | |
| Regex2026.04 | — | — | — | 92.5 |