Truthful Question Answering on TruthfulQA (TruthRate and InfoRate)
83TruthRateSEA
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SEABase Model=LLaMA2-13B-Chat2025.05 | 83 | 100 | |
| BoN64Base Model=LLaMA2-13B-Chat2025.05 | 76 | 100 | |
| SFTBase Model=LLaMA2-13B-Chat2025.05 | 73 | 100 | |
| ITIBase Model=LLaMA2-7B-Chat2026.06 | 57.64 | 27.84 | |
| TruthProbeBase LLMBase Model=LLaMA2-7B-Chat2026.06 | 56.91 | 27 | |
| LLaMA2-7B-ChatMode=Vanilla2026.06 | 56.4 | 25.56 | |
| TruthProbeFT LLMBase Model=LLaMA2-7B-Chat2026.06 | 55.38 | 29.02 | |
| L3-CATBackbone=Llama 32026.05 | 43 | — | |
| LPA (ours)Backbone=Llama 32026.05 | 38.1 | 57.4 | |
| Base (8B Instruct)Backbone=Llama 32026.05 | 37 | 52.4 | |
| L2-CATBackbone=Llama 22026.05 | 35.3 | — | |
| LPA-overfit (ours)Backbone=Llama 32026.05 | 33.8 | 54.6 | |
| L3-LATBackbone=Llama 32026.05 | 33.2 | 56.6 | |
| LPA (ours)Backbone=Llama 22026.05 | 31.5 | 46.4 | |
| Base (7B chat-hf)Backbone=Llama 22026.05 | 30.7 | 46.2 | |
| LPA-overfit (ours)Backbone=Llama 22026.05 | 30.6 | 46.1 | |
| L2-LATBackbone=Llama 22026.05 | 28.6 | 47.8 |