Truthfulness Evaluation on TruthfulQA (BLEU-Acc)
54.7BLEU-AccSG
Evaluation Results
| Method | Links | |
|---|---|---|
| SGBackbone=GPT-OSS-20B, Candidate Generation=k=10 short answers, Decision Layer=smooth (sigmoid)2025.10 | 54.7 | |
| ER-DBackbone=GPT-OSS-20B, Candidate Generation=k=10 short answers2025.10 | 52.1 | |
| DBackbone=GPT-OSS-20B, Candidate Generation=k=10 short answers2025.10 | 52 | |
| ER-GBackbone=GPT-OSS-20B, Candidate Generation=k=10 short answers2025.10 | 52 | |
| SCBackbone=GPT-OSS-20B, Candidate Generation=k=10 short answers2025.10 | 52 | |
| SGBackbone=Llama-3.2-1B, Candidate Generation=k=10 short answers, Decision Layer=smooth (sigmoid)2025.10 | 51.4 | |
| MIBackbone=GPT-OSS-20B, Candidate Generation=k=10 short answers2025.10 | 51.4 | |
| GBackbone=GPT-OSS-20B, Candidate Generation=k=10 short answers2025.10 | 51.2 | |
| SCBackbone=LLaMA-2-13B, Candidate Generation=k=10 short answers2025.10 | 51.1 | |
| SGBackbone=LLaMA-2-13B, Candidate Generation=k=10 short answers, Decision Layer=smooth (sigmoid)2025.10 | 50.9 | |
| SGBackbone=LLaMA-2-7B, Candidate Generation=k=10 short answers, Decision Layer=smooth (sigmoid)2025.10 | 50.4 | |
| SGBackbone=Llama-3.1-8B, Candidate Generation=k=10 short answers, Decision Layer=smooth (sigmoid)2025.10 | 50.3 | |
| GBackbone=LLaMA-2-7B, Candidate Generation=k=10 short answers2025.10 | 50.1 | |
| MIBackbone=LLaMA-2-7B, Candidate Generation=k=10 short answers2025.10 | 49.8 | |
| ER-GBackbone=LLaMA-2-13B, Candidate Generation=k=10 short answers2025.10 | 48.8 | |
| MIBackbone=LLaMA-2-13B, Candidate Generation=k=10 short answers2025.10 | 48.6 | |
| GBackbone=LLaMA-2-13B, Candidate Generation=k=10 short answers2025.10 | 48.5 | |
| MIBackbone=Llama-3.1-8B, Candidate Generation=k=10 short answers2025.10 | 47.8 | |
| GBackbone=Llama-3.1-8B, Candidate Generation=k=10 short answers2025.10 | 47.8 | |
| ER-DBackbone=LLaMA-2-13B, Candidate Generation=k=10 short answers2025.10 | 46.7 | |
| SCBackbone=Llama-3.2-1B, Candidate Generation=k=10 short answers2025.10 | 46.6 | |
| GBackbone=Llama-3.2-1B, Candidate Generation=k=10 short answers2025.10 | 46.2 | |
| SCBackbone=Llama-3.1-8B, Candidate Generation=k=10 short answers2025.10 | 46 | |
| MIBackbone=Llama-3.2-1B, Candidate Generation=k=10 short answers2025.10 | 46 | |
| DBackbone=Llama-3.2-1B, Candidate Generation=k=10 short answers2025.10 | 45.6 | |
| SCBackbone=LLaMA-2-7B, Candidate Generation=k=10 short answers2025.10 | 45.2 | |
| ER-GBackbone=Llama-3.2-1B, Candidate Generation=k=10 short answers2025.10 | 45 | |
| ER-DBackbone=Llama-3.1-8B, Candidate Generation=k=10 short answers2025.10 | 44.7 | |
| DBackbone=Llama-3.1-8B, Candidate Generation=k=10 short answers2025.10 | 44.5 | |
| ER-GBackbone=LLaMA-2-7B, Candidate Generation=k=10 short answers2025.10 | 44.4 | |
| ER-GBackbone=Llama-3.1-8B, Candidate Generation=k=10 short answers2025.10 | 44.4 | |
| ER-DBackbone=Llama-3.2-1B, Candidate Generation=k=10 short answers2025.10 | 44.4 | |
| DBackbone=LLaMA-2-7B, Candidate Generation=k=10 short answers2025.10 | 44 | |
| ER-DBackbone=LLaMA-2-7B, Candidate Generation=k=10 short answers2025.10 | 43.5 | |
| DBackbone=LLaMA-2-13B, Candidate Generation=k=10 short answers2025.10 | 37 |