List your company
ResearchFormerly WizWand

Hallucination detection on MMLU-Pro

87.08AUROCDRIFT
48.516858.528468.5478.5516Jan 20, 2026Feb 14, 2026Mar 11, 2026Apr 5, 2026Apr 30, 2026May 25, 2026Jun 20, 2026
Updated 4mo ago

Evaluation results

MethodLinks
DRIFTModel=Qwen-2.5-7B-Instruct, Input=answer2026.0187.08——
DRIFTModel=Qwen-2.5-7B-Instruct, Input=question2026.0182.21——
HaloScopeModel=Qwen-2.5-7B-Instruct, Input=question2026.0181.08——
HaloScopeModel=LLaMA 2 Chat 7B, Input=question2026.0179.56——
DRIFTModel=LLaMA 2 Chat 7B, Input=answer2026.0177.85——
HaloScopeModel=Gemma-3-4b-it, Input=question2026.0177.73——
HaloScopeModel=Qwen-2.5-7B-Instruct, Input=answer2026.0177.56——
HaloScopeModel=Gemma-3-4b-it, Input=answer2026.0177.56——
Semantic EntropyModel=LLaMA 2 Chat 7B2026.0176.76——
DRIFTModel=Gemma-3-4b-it, Input=answer2026.0175.85——
Attention probeModel=Qwen2.5-7B, Generation Stage=Post-generation2026.0675.42——
HaloScopeModel=LLaMA 2 Chat 7B, Input=answer2026.0174.83——
DRIFTModel=Gemma-3-4b-it, Input=question2026.0174.65——
Attention probeModel=Qwen2.5-7B, Generation Stage=Pre-generation2026.0674.4——
Semantic EntropyModel=Qwen-2.5-7B-Instruct2026.0174.05——
EntropyModel=Qwen2.5-7B, Generation Stage=Post-generation2026.0672.32——
DRIFTModel=LLaMA 2 Chat 7B, Input=question2026.0171.22——
Linear probeModel=Qwen2.5-7B, Generation Stage=Pre-generation2026.0669.23——
Linear probeModel=Qwen2.5-7B, Generation Stage=Post-generation2026.0669.18——
EntropyModel=Qwen2.5-3B, Generation Stage=Post-generation2026.0664.99——
Linear probeModel=Qwen2.5-3B, Generation Stage=Post-generation2026.0664.03——
Linear probeModel=Qwen2.5-3B, Generation Stage=Pre-generation2026.0663.89——
Self-check, zero-shotModel=Qwen2.5-7B, Generation Stage=Pre-generation2026.0660.2——
Question lengthModel=Qwen2.5-7B, Generation Stage=Pre-generation2026.0660.15——
Question lengthModel=Qwen2.5-3B, Generation Stage=Pre-generation2026.0659.21——
Semantic EntropyModel=Gemma-3-4b-it2026.0157.92——
Attention probe, soft targetModel=Qwen2.5-7B, Generation Stage=Pre-generation2026.0656.95——
Self-check, zero-shotModel=Qwen2.5-3B, Generation Stage=Pre-generation2026.0653.73——
Attention probe, soft targetModel=Qwen2.5-3B, Generation Stage=Pre-generation2026.0653.55——
Attention probeModel=Qwen2.5-3B, Generation Stage=Post-generation2026.0650——
Attention probeModel=Qwen2.5-3B, Generation Stage=Pre-generation2026.0650——
FACTOOLModels=Qwen2.5-7B2025.11—59.669.23
FACTOOLModels=Llama3.1-8B2025.11—55.3368.84
FACTOOLModels=Mistral-8B2025.11—62.2671.43
FIREModels=Qwen2.5-7B2025.11—61.0569.23
FIREModels=Llama3.1-8B2025.11—58.8569.51
FIREModels=Mistral-8B2025.11—58.2371.11
HaluAgentModels=Qwen2.5-7B2025.11—54.6855
HaluAgentModels=Llama3.1-8B2025.11—54.4154.05
HaluAgentModels=Mistral-8B2025.11—54.2970.37
LEAPModels=Qwen2.5-7B2025.11—69.8175.31
LEAPModels=Llama3.1-8B2025.11—64.2371.18
LEAPModels=Mistral-8B2025.11—63.2171.15
LN-entropyModels=Qwen2.5-7B2025.11—56.6767.5
LN-entropyModels=Llama3.1-8B2025.11—58.3468.51
LN-entropyModels=Mistral-8B2025.11—58.6768.04
PerplexityModels=Qwen2.5-7B2025.11—57.3368
PerplexityModels=Llama3.1-8B2025.11—5570.97
PerplexityModels=Mistral-8B2025.11—56.6767.17
SAFEModels=Qwen2.5-7B2025.11—55.8168.28
SAFEModels=Llama3.1-8B2025.11—57.9670.49
SAFEModels=Mistral-8B2025.11—60.6172.19
Self-Check(0)Models=Qwen2.5-7B2025.11—56.6669.54
Self-Check(0)Models=Llama3.1-8B2025.11—54.8870.74
Self-Check(0)Models=Mistral-8B2025.11—60.1470.37
Self-Check(3)Models=Qwen2.5-7B2025.11—59.5664.52
Self-Check(3)Models=Llama3.1-8B2025.11—55.2570.67
Self-Check(3)Models=Mistral-8B2025.11—59.1270.42
Semantic EntropyModels=Qwen2.5-7B2025.11—55.6766.67
Semantic EntropyModels=Llama3.1-8B2025.11—55.3365.59
Semantic EntropyModels=Mistral-8B2025.11—54.6769.64