List your company
ResearchFormerly WizWand

Hallucination Detection on MATH 500

88.89AUROCEigenWD M1
21.269238.824656.3873.9354Jan 24, 2026Feb 17, 2026Mar 13, 2026Apr 6, 2026Apr 30, 2026May 24, 2026Jun 17, 2026
Updated 4mo ago

Evaluation results

MethodLinks
EigenWD M1Target Model=DeepSeek-Chat, Teacher Model=Llama-2-7B, Detection Setting=Black-box case study2026.0388.89
ARS (CCS)Model=DeepSeek-R1-Distill-Llama-8B, Judge Method=ROUGE2026.0188
ERTarget Model=DeepSeek-Chat, Teacher Model=Llama-2-7B, Detection Setting=Black-box case study2026.0387.6
ARS (CCS)Model=DeepSeek-R1-Distill-Llama-8B, Single Sampling=true, Supervision=false2026.0186.38
ARS (CCS)Model=DeepSeek-R1-Distill-Llama-8B, Judge Method=DeepSeek-R1-Distill-Qwen-32B2026.0186.24
TSVModel=DeepSeek-R1-Distill-Llama-8B, Judge Method=ROUGE2026.0184.71
LSTarget Model=DeepSeek-Chat, Teacher Model=Llama-2-7B, Detection Setting=Black-box case study2026.0383.35
TSVModel=DeepSeek-R1-Distill-Llama-8B, Judge Method=BLEURT2026.0181.36
ARS (CCS)Model=DeepSeek-R1-Distill-Llama-8B, Judge Method=BLEURT2026.0181
ARS (Probing)Model=DeepSeek-R1-Distill-Llama-8B, Single Sampling=true, Supervision=true2026.0179.95
ARS (CCS)Model=Qwen3-8B, Single Sampling=true, Supervision=false2026.0178.66
TSVModel=DeepSeek-R1-Distill-Llama-8B, Judge Method=DeepSeek-R1-Distill-Qwen-32B2026.0178.64
ARS (Probing)Model=Qwen3-8B, Single Sampling=true, Supervision=true2026.0178.17
G-DetectorModel=DeepSeek-R1-Distill-Llama-8B, Judge Method=ROUGE2026.0171.48
FESSignal=Free-energy + SFF2026.0668.3
Best spec.Description=Strongest non-FES in-tree attention-spectral baseline2026.0667.4
G-DetectorModel=DeepSeek-R1-Distill-Llama-8B, Single Sampling=true, Supervision=true2026.0164.45
TSVModel=DeepSeek-R1-Distill-Llama-8B, Single Sampling=true, Supervision=true2026.0163.24
TSVModel=Qwen3-8B, Single Sampling=true, Supervision=true2026.0163.12
RACEModel=Qwen3-8B, Single Sampling=false, Supervision=false2026.0163.02
G-DetectorModel=DeepSeek-R1-Distill-Llama-8B, Judge Method=BLEURT2026.0161.44
DSETarget Model=DeepSeek-Chat, Teacher Model=Llama-2-7B, Detection Setting=Black-box case study2026.0359.97
Best LMDescription=Best language-model probability baseline among MSP and PPL−12026.0659.6
SelfCKGPTModel=DeepSeek-R1-Distill-Llama-8B, Single Sampling=false, Supervision=false2026.0159.15
Best rep.Description=Strongest reported reference hidden-state or sampling baseline2026.0658
G-DetectorModel=Qwen3-8B, Single Sampling=true, Supervision=true2026.0157.67
RHDModel=DeepSeek-R1-Distill-Llama-8B, Single Sampling=true, Supervision=false2026.0156.5
Semantic EntropyModel=Qwen3-8B, Single Sampling=false, Supervision=false2026.0156.13
ESTarget Model=DeepSeek-Chat, Teacher Model=Llama-2-7B, Detection Setting=Black-box case study2026.0355.87
SelfCKGPTModel=Qwen3-8B, Single Sampling=false, Supervision=false2026.0155.47
RACEModel=DeepSeek-R1-Distill-Llama-8B, Single Sampling=false, Supervision=false2026.0153.55
LNETarget Model=DeepSeek-Chat, Teacher Model=Llama-2-7B, Detection Setting=Black-box case study2026.0353.44
PerplexityModel=Qwen3-8B, Single Sampling=true, Supervision=false2026.0151.62
RHDModel=Qwen3-8B, Single Sampling=true, Supervision=false2026.0150.51
Lexical SimilarityModel=DeepSeek-R1-Distill-Llama-8B, Single Sampling=false, Supervision=false2026.0149.92
Verbalized CertaintyModel=DeepSeek-R1-Distill-Llama-8B, Single Sampling=true, Supervision=false2026.0148.98
Lexical SimilarityModel=Qwen3-8B, Single Sampling=false, Supervision=false2026.0144.13
Semantic EntropyModel=DeepSeek-R1-Distill-Llama-8B, Single Sampling=false, Supervision=false2026.0143.6
PerplexityModel=DeepSeek-R1-Distill-Llama-8B, Single Sampling=true, Supervision=false2026.0140.96
G-DetectorModel=DeepSeek-R1-Distill-Llama-8B, Judge Method=DeepSeek-R1-Distill-Qwen-32B2026.0138.06
Verbalized CertaintyModel=Qwen3-8B, Single Sampling=true, Supervision=false2026.0123.87