Multiple-choice Question Answering on TruthfulQA Multiple-choice
57.77MC1 ScoreCD
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| CDBackbone=Llama-3-70B2026.06 | 57.77 | 76.65 | 47.08 | |
| DeCoRe (static)Backbone=Llama-3-70B2026.06 | 51.29 | 72.02 | 40.24 | |
| DoLa (high)Backbone=Llama-3-70B2026.06 | 49.69 | 70.88 | 38.01 | |
| GreedyBackbone=Llama-3-70B2026.06 | 49.57 | 70.6 | 37.85 | |
| DoLa (low)Backbone=Llama-3-70B2026.06 | 49.45 | 70.58 | 37.75 | |
| ITIBackbone=Llama-3-70B2026.06 | 48.96 | 67.04 | 37.27 | |
| DCOBackbone=Llama-3-70B2026.06 | 48.23 | 69.56 | 36.73 | |
| SafeLMsetup=full2026.04 | 46.1 | 61.2 | — | |
| RAdArchitecture=Zephyr-7B, Concept Direction=truthfulness2026.01 | 44.9 | 62.3 | — | |
| ITIBackbone=Llama-3-8B2026.06 | 43.7 | 62.78 | 34.91 | |
| ADBackbone=Llama-3-70B2026.06 | 42.23 | 67.56 | 35.37 | |
| DCOBackbone=Llama-3-8B2026.06 | 40.39 | 58.68 | 32.21 | |
| GreedyBackbone=Llama-3-8B2026.06 | 39.41 | 55.69 | 30.31 | |
| RAG only2026.04 | 39.1 | 54.7 | — | |
| DoLa (low)Backbone=Llama-3-8B2026.06 | 39.05 | 55.65 | 30.06 | |
| Base modelArchitecture=Zephyr-7B2026.01 | 39 | 55 | — | |
| DoLa (high)Backbone=Llama-3-8B2026.06 | 38.68 | 55.64 | 30.19 | |
| DeCoRe (static)Backbone=Llama-3-8B2026.06 | 38.68 | 55.74 | 29.8 | |
| RAbArchitecture=Zephyr-7B, Concept Direction=random2026.01 | 38.6 | 55.6 | — | |
| RAdArchitecture=Zephyr-7B, Concept Direction=random2026.01 | 38.4 | 55.9 | — | |
| SafeLMsetup=no MG2026.04 | 37.9 | 53.4 | — | |
| LLM-CASBackbone=LLaMA2-7B-CHAT2025.12 | 35.47 | 51.45 | 26.43 | |
| SADIBackbone=LLaMA2-7B-CHAT2025.12 | 34.88 | 52.5 | 25.79 | |
| ITIBackbone=LLaMA2-7B-CHAT2025.12 | 34.64 | 51.55 | 25.32 | |
| RAdArchitecture=Mistral-7B, Concept Direction=truthfulness2026.01 | 34.1 | 49.9 | — | |
| CAABackbone=LLaMA2-7B-CHAT2025.12 | 34.03 | 52.76 | 25.62 | |
| BaselineBackbone=LLaMA2-7B-CHAT2025.12 | 33.41 | 51.07 | 24.76 | |
| ADBackbone=Llama-3-8B2026.06 | 31.21 | 55.3 | 28.28 | |
| Vanilla fine-tune2026.04 | 31.2 | 48.1 | — | |
| RAdArchitecture=Mistral-7B, Concept Direction=random2026.01 | 28.6 | 42.9 | — | |
| RAbArchitecture=Mistral-7B, Concept Direction=random2026.01 | 28.4 | 43.2 | — | |
| Base modelArchitecture=Mistral-7B2026.01 | 28.2 | 42.6 | — | |
| RAbArchitecture=Zephyr-7B, Concept Direction=truthfulness2026.01 | 26.1 | 40 | — | |
| RAbArchitecture=Mistral-7B, Concept Direction=truthfulness2026.01 | 24.1 | 37.4 | — |