Question Answering on TruthfulQA (MC1, MC2, MC3, Avg. Δ)
68.25MC2DECOVEC
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DECOVECModel=Llama-3-8B, Protocol=Few-Shot (Rand. Ext.)2026.04 | 68.25 | 49.93 | 39.86 | 1.16 | |
| DECOVECModel=Llama-3-8B, Protocol=Few-Shot (KATE)2026.04 | 67.53 | 48.1 | 38.89 | 1.68 | |
| DECOVECModel=Llama-3-8B, Protocol=Few-Shot (BM25)2026.04 | 67.1 | 48.52 | 38.61 | 0.82 | |
| DECOVECModel=Llama-3-8B, Protocol=Few-Shot (Rand. Std.)2026.04 | 65.76 | 44.87 | 36.87 | 1.96 | |
| DECOVECModel=Gemma-2-9B, Protocol=Few-Shot (KATE)2026.04 | 63.28 | 46.55 | 34.17 | 1.06 | |
| GRM + CIRMBackbone=Gemma-2-9B-it2026.04 | 63.06 | — | — | — | |
| DECOVECModel=Gemma-2-9B, Protocol=Few-Shot (Rand. Ext.)2026.04 | 63.04 | 45.43 | 33.64 | 0.78 | |
| GRMBackbone=Gemma-2-9B-it2026.04 | 62.78 | — | — | — | |
| DECOVECModel=Gemma-2-9B, Protocol=Few-Shot (BM25)2026.04 | 60.93 | 43.32 | 32 | 1.15 | |
| DECOVECModel=Qwen2-7B, Protocol=Few-Shot (KATE)2026.04 | 60.56 | 43.32 | 32.69 | 1.35 | |
| FsfairX + CIRMBackbone=Gemma-2-9B-it2026.04 | 60.39 | — | — | — | |
| DECOVECModel=Qwen2-1.5B, Protocol=Few-Shot (Rand. Ext.)2026.04 | 60.33 | 42.76 | 32.24 | 1.81 | |
| FsfairXBackbone=Gemma-2-9B-it2026.04 | 59.99 | — | — | — | |
| GRM + CIRMBackbone=Llama-3-8B-Instruct2026.04 | 59.5 | — | — | — | |
| DECOVECModel=Qwen2-7B, Protocol=Few-Shot (Rand. Ext.)2026.04 | 59.43 | 42.33 | 31.62 | 1.43 | |
| DECOVECModel=Qwen2-7B, Protocol=Few-Shot (BM25)2026.04 | 59.34 | 42.48 | 31.3 | 1.1 | |
| FsfairXBackbone=Llama-3-8B-Instruct2026.04 | 59.21 | — | — | — | |
| FsfairX + CIRMBackbone=Llama-3-8B-Instruct2026.04 | 58.89 | — | — | — | |
| GRMBackbone=Llama-3-8B-Instruct2026.04 | 58.5 | — | — | — | |
| Zero-ShotModel=Llama-3-8B, Protocol=Zero-Shot2026.04 | 58.06 | 37.83 | 30.69 | — | |
| DECOVECModel=Llama-2-7B, Protocol=Few-Shot (Rand. Ext.)2026.04 | 57.36 | 39.66 | 29.32 | 2.19 | |
| DECOVECModel=Llama-2-7B, Protocol=Few-Shot (BM25)2026.04 | 57.12 | 39.38 | 28.66 | 2.61 | |
| DECOVECModel=Llama-2-7B, Protocol=Few-Shot (KATE)2026.04 | 57.06 | 39.94 | 29.16 | 2.38 | |
| DECOVECModel=Qwen2-7B, Protocol=Few-Shot (Rand. Std.)2026.04 | 56.39 | 38.54 | 29.94 | 1.98 | |
| DECOVECModel=Gemma-2-9B, Protocol=Few-Shot (Rand. Std.)2026.04 | 56.11 | 39.24 | 28.55 | 1.88 | |
| DECOVECModel=Yi-6B, Protocol=Few-Shot (KATE)2026.04 | 55.66 | 38.68 | 28.68 | 2.79 | |
| DECOVECModel=Yi-6B, Protocol=Few-Shot (Rand. Ext.)2026.04 | 55.14 | 38.82 | 28.51 | 2.64 | |
| DECOVECModel=Llama-2-7B, Protocol=Few-Shot (Rand. Std.)2026.04 | 54.97 | 36.01 | 27.92 | 2.05 | |
| DECOVECModel=Yi-6B, Protocol=Few-Shot (BM25)2026.04 | 54.43 | 38.68 | 27.85 | 2.63 | |
| DECOVECModel=Qwen2-1.5B, Protocol=Few-Shot (KATE)2026.04 | 53.89 | 37.41 | 27.47 | 0.82 | |
| DECOVECModel=Qwen2-1.5B, Protocol=Few-Shot (BM25)2026.04 | 53.84 | 36.71 | 26.72 | 1.51 | |
| DECOVECModel=Yi-6B, Protocol=Few-Shot (Rand. Std.)2026.04 | 52.01 | 34.18 | 26.67 | 1.59 | |
| DECOVECModel=Qwen2-0.5B, Protocol=Few-Shot (KATE)2026.04 | 50.43 | 34.6 | 24.98 | 1.37 | |
| Zero-ShotModel=Qwen2-7B, Protocol=Zero-Shot2026.04 | 50.13 | 32.49 | 26.25 | — | |
| DECOVECModel=Qwen2-1.5B, Protocol=Few-Shot (Rand. Std.)2026.04 | 49.91 | 33.47 | 25.02 | 0.64 | |
| DECOVECModel=Qwen2-0.5B, Protocol=Few-Shot (Rand. Ext.)2026.04 | 48.01 | 32.21 | 23.73 | 1.56 | |
| DECOVECModel=Qwen2-0.5B, Protocol=Few-Shot (BM25)2026.04 | 47.83 | 32.35 | 23.09 | 1.43 | |
| DECOVECModel=Qwen2-0.5B, Protocol=Few-Shot (Rand. Std.)2026.04 | 47.54 | 32.35 | 23.46 | 1.95 | |
| Zero-ShotModel=Gemma-2-9B, Protocol=Zero-Shot2026.04 | 45.06 | 27 | 22.08 | — | |
| Zero-ShotModel=Llama-2-7B, Protocol=Zero-Shot2026.04 | 44.94 | 26.58 | 21.76 | — | |
| Zero-ShotModel=Qwen2-1.5B, Protocol=Zero-Shot2026.04 | 44.7 | 26.3 | 20.78 | — | |
| Zero-ShotModel=Yi-6B, Protocol=Zero-Shot2026.04 | 44.64 | 25.46 | 21.5 | — | |
| Zero-ShotModel=Qwen2-0.5B, Protocol=Zero-Shot2026.04 | 40.78 | 25.18 | 19.55 | — | |
| KICK-SETTLEDefense=SDD, Model=Llama-2-7B, Attack Type=Kick-settle2026.05 | — | — | — | 0.042 | |
| KICK-SETTLEDefense=SDD, Model=Qwen3-8B, Attack Type=Kick-settle2026.05 | — | — | — | 0.015 | |
| KICK-SETTLEDefense=SDD, Model=Llama-3.1-8B, Attack Type=Kick-settle2026.05 | — | — | — | 0.063 |