Hallucination Detection on TruthfulQA (AUROC and F1-Score)
0.8851AUROCCausalGaze
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CausalGazeLLM=Mistral-7B2026.04 | 0.8851 | 79.71 | |
| CausalGazeLLM=Llama2-7B2026.04 | 0.8805 | 81.89 | |
| HaluGNNLLM=Llama2-7B2026.04 | 0.8803 | 77.14 | |
| CausalGazeLLM=Qwen2-7B2026.04 | 0.868 | 80.71 | |
| HaluGNNLLM=Mistral-7B2026.04 | 0.8647 | 77.47 | |
| HaluGNNLLM=Qwen2-7B2026.04 | 0.8392 | 77.06 | |
| SAPLMALLM=Qwen2-7B2026.04 | 0.817 | 82.8 | |
| ICR-ProbeLLM=Llama2-7B2026.04 | 0.8142 | 78.58 | |
| SAPLMALLM=Mistral-7B2026.04 | 0.8112 | 78.83 | |
| PCNETModel=Qwen3-4B2026.05 | 0.81 | 0.78 | |
| ICR-ProbeLLM=Mistral-7B2026.04 | 0.7993 | 80 | |
| ICR-ProbeLLM=Qwen2-7B2026.04 | 0.7937 | 77.4 | |
| PCNETModel=Mistral-7B-v0.32026.05 | 0.79 | 0.76 | |
| SAPLMALLM=Llama2-7B2026.04 | 0.782 | 79.03 | |
| FLaG (Full)Backbone=LLaMA-3-8B-Instruct2026.05 | 0.7576 | — | |
| PCNETModel=Llama-3.1-8B2026.05 | 0.75 | 0.72 | |
| FLaG (Full)Backbone=Qwen2.5-7B-Instruct2026.05 | 0.7304 | — | |
| FLaG (Semi)Backbone=LLaMA-3-8B-Instruct2026.05 | 0.7249 | — | |
| SAPLMABackbone=Qwen2.5-7B-Instruct2026.05 | 0.7192 | — | |
| SAPLMABackbone=LLaMA-3-8B-Instruct2026.05 | 0.7168 | — | |
| TTPDBackbone=Qwen2.5-7B-Instruct2026.05 | 0.711 | — | |
| Linear probeBackbone=LLaMA-3-8B-Instruct2026.05 | 0.7073 | — | |
| FLaG (Semi)Backbone=Qwen2.5-7B-Instruct2026.05 | 0.7059 | — | |
| Probe-LRBackbone=Qwen2.5-7B-Instruct2026.05 | 0.7036 | — | |
| HaloScopeBackbone=Qwen2.5-7B-Instruct2026.05 | 0.6966 | — | |
| TTPDBackbone=LLaMA-3-8B-Instruct2026.05 | 0.6827 | — | |
| Linear probeBackbone=Qwen2.5-7B-Instruct2026.05 | 0.6801 | — | |
| LLM-CheckLLM=Mistral-7B2026.04 | 0.6732 | 65.4 | |
| HaloScopeBackbone=LLaMA-3-8B-Instruct2026.05 | 0.6679 | — | |
| LN-EntropyLLM=Qwen2-7B2026.04 | 0.667 | 65.94 | |
| EarlyDetecBackbone=Qwen2.5-7B-Instruct2026.05 | 0.6615 | — | |
| Semantic EntropyLLM=Qwen2-7B2026.04 | 0.661 | 67.87 | |
| EarlyDetecBackbone=LLaMA-3-8B-Instruct2026.05 | 0.6601 | — | |
| PCNETModel=Llama-3.2-1B2026.05 | 0.66 | 0.69 | |
| Semantic EntropyLLM=Mistral-7B2026.04 | 0.6557 | 61.36 | |
| Probe-LRBackbone=LLaMA-3-8B-Instruct2026.05 | 0.6541 | — | |
| PerplexityLLM=Qwen2-7B2026.04 | 0.651 | 66.3 | |
| SelfCKGPTBackbone=Qwen2.5-7B-Instruct2026.05 | 0.647 | — | |
| SPUQBackbone=LLaMA-3-8B-Instruct2026.05 | 0.6446 | — | |
| TSVBackbone=LLaMA-3-8B-Instruct2026.05 | 0.6394 | — | |
| P(True)LLM=Qwen2-7B2026.04 | 0.637 | 65.42 | |
| LLM-CheckLLM=Qwen2-7B2026.04 | 0.6316 | 60.06 | |
| EGHBackbone=LLaMA-3-8B-Instruct2026.05 | 0.6289 | — | |
| Semantic EntropyBackbone=Qwen2.5-7B-Instruct2026.05 | 0.6251 | — | |
| VerbalizeBackbone=LLaMA-3-8B-Instruct2026.05 | 0.6233 | — | |
| Semantic EntropyLLM=Llama2-7B2026.04 | 0.6217 | 61.45 | |
| SelfCheckGPTLLM=Qwen2-7B2026.04 | 0.617 | 62.2 | |
| LLM-CheckLLM=Llama2-7B2026.04 | 0.616 | 59.26 | |
| EGHBackbone=Qwen2.5-7B-Instruct2026.05 | 0.6158 | — | |
| LN-EntropyLLM=Llama2-7B2026.04 | 0.6151 | 61.87 | |
| PerplexityBackbone=LLaMA-3-8B-Instruct2026.05 | 0.6113 | — | |
| HaloScopeModel=Llama-3.2-1B2026.05 | 0.61 | 0.67 | |
| EigenScoreLLM=Mistral-7B2026.04 | 0.6012 | 58.6 | |
| SelfCKGPTBackbone=LLaMA-3-8B-Instruct2026.05 | 0.5966 | — | |
| SPUQBackbone=Qwen2.5-7B-Instruct2026.05 | 0.5912 | — | |
| HaloScopeModel=Llama-3.1-8B2026.05 | 0.59 | 0.67 | |
| Lexical SimilarityBackbone=Qwen2.5-7B-Instruct2026.05 | 0.5833 | — | |
| Semantic EntropyBackbone=LLaMA-3-8B-Instruct2026.05 | 0.5808 | — | |
| TSVBackbone=Qwen2.5-7B-Instruct2026.05 | 0.5796 | — | |
| SelfCheckGPTLLM=Mistral-7B2026.04 | 0.5771 | 53.45 | |
| LN-EntropyLLM=Mistral-7B2026.04 | 0.5763 | 57.2 | |
| EigenScoreBackbone=LLaMA-3-8B-Instruct2026.05 | 0.5712 | — | |
| PerplexityLLM=Llama2-7B2026.04 | 0.5677 | 60.34 | |
| PerplexityBackbone=Qwen2.5-7B-Instruct2026.05 | 0.564 | — | |
| PerplexityLLM=Mistral-7B2026.04 | 0.5538 | 54.21 | |
| VerbalizeBackbone=Qwen2.5-7B-Instruct2026.05 | 0.5507 | — | |
| HaloScopeModel=Mistral-7B-v0.32026.05 | 0.55 | 0.67 | |
| SEPModel=Mistral-7B-v0.32026.05 | 0.54 | 0.67 | |
| SEPModel=Qwen3-4B2026.05 | 0.54 | 0.67 | |
| EigenScoreLLM=Qwen2-7B2026.04 | 0.537 | 54.89 | |
| EigenScoreBackbone=Qwen2.5-7B-Instruct2026.05 | 0.5324 | — | |
| SEPModel=Llama-3.2-1B2026.05 | 0.53 | 0.67 | |
| Token NLLModel=Qwen3-4B2026.05 | 0.53 | 0.67 | |
| HaloScopeModel=Qwen3-4B2026.05 | 0.53 | 0.68 | |
| SelfCheckGPTLLM=Llama2-7B2026.04 | 0.5295 | 53.45 | |
| CCSBackbone=Qwen2.5-7B-Instruct2026.05 | 0.529 | — | |
| P(True)LLM=Mistral-7B2026.04 | 0.526 | 51.25 | |
| CCSBackbone=LLaMA-3-8B-Instruct2026.05 | 0.5219 | — | |
| Self-evaluationBackbone=LLaMA-3-8B-Instruct2026.05 | 0.5203 | — | |
| Token NLLModel=Mistral-7B-v0.32026.05 | 0.52 | 0.67 | |
| EigenScoreLLM=Llama2-7B2026.04 | 0.5193 | 52.86 | |
| P(True)LLM=Llama2-7B2026.04 | 0.5181 | 55.56 | |
| Lexical SimilarityBackbone=LLaMA-3-8B-Instruct2026.05 | 0.5154 | — | |
| Self-evaluationBackbone=Qwen2.5-7B-Instruct2026.05 | 0.5084 | — | |
| Token NLLModel=Llama-3.1-8B2026.05 | 0.49 | 0.67 | |
| Token NLLModel=Llama-3.2-1B2026.05 | 0.47 | 0.67 | |
| SEPModel=Llama-3.1-8B2026.05 | 0.43 | 0.67 | |
| AutoFactModel=Llama-3.1-8B2026.05 | 0.35 | 0.67 | |
| AutoFactModel=Llama-3.2-1B2026.05 | 0.34 | 0.67 | |
| AutoFactModel=Mistral-7B-v0.32026.05 | 0.34 | 0.67 | |
| AutoFactModel=Qwen3-4B2026.05 | 0.34 | 0.67 |