List your company
ResearchFormerly WizWand

Hallucination Detection on RAGTruth

0.8535AUROCSpikeScore
0.371460.4966050.621750.746895Jan 27, 2026Feb 20, 2026Mar 16, 2026Apr 9, 2026May 3, 2026May 27, 2026Jun 20, 2026
Updated 4mo ago

Evaluation results

MethodLinks
SpikeScoreLLM Backbone=Qwen3-14B, Method Category=Cross-domain specialized methods2026.010.8535
SpikeScoreLLM Backbone=Qwen3-8B, Method Category=Cross-domain specialized methods2026.010.8291
SpikeScoreLLM Backbone=Llama3.1-8B, Method Category=Cross-domain specialized methods2026.010.8154
D: Act. VarianceBackbone=Llama-3.3-70B2026.050.78
ICR ProbeLLM Backbone=Qwen3-14B, Method Category=Cross-domain specialized methods2026.010.7653
SpikeScoreLLM Backbone=Llama3.2-3B, Method Category=Cross-domain specialized methods2026.010.7631
ICR ProbeLLM Backbone=Qwen3-8B, Method Category=Cross-domain specialized methods2026.010.7426
ICR ProbeLLM Backbone=Llama3.2-3B, Method Category=Cross-domain specialized methods2026.010.7416
ICR ProbeLLM Backbone=Llama3.1-8B, Method Category=Cross-domain specialized methods2026.010.7405
EigenScoreLLM Backbone=Llama3.1-8B, Method Category=Training-free methods2026.010.6974
Lexical SimilarityLLM Backbone=Llama3.1-8B, Method Category=Training-free methods2026.010.6958
EigenScoreLLM Backbone=Qwen3-14B, Method Category=Training-free methods2026.010.6904
Semantic EntropyLLM Backbone=Qwen3-14B, Method Category=Training-free methods2026.010.6796
Attention probeModel=Llama-2-13B, Generation Phase=Pre-generation2026.060.6734
EntropyModel=Mistral-7B, Generation Phase=Post-generation2026.060.6671
Attention probeModel=Mistral-7B, Generation Phase=Post-generation2026.060.6531
Attention probeModel=Mistral-7B, Generation Phase=Pre-generation2026.060.6524
Lexical SimilarityLLM Backbone=Qwen3-14B, Method Category=Training-free methods2026.010.6479
Lexical SimilarityLLM Backbone=Qwen3-8B, Method Category=Training-free methods2026.010.644
PerplexityLLM Backbone=Qwen3-14B, Method Category=Training-free methods2026.010.642
Semantic EntropyLLM Backbone=Qwen3-8B, Method Category=Training-free methods2026.010.6388
SAPLMALLM Backbone=Qwen3-14B, Method Category=Training-based methods2026.010.6338
EigenScoreLLM Backbone=Llama3.2-3B, Method Category=Training-free methods2026.010.6329
EigenScoreLLM Backbone=Qwen3-8B, Method Category=Training-free methods2026.010.6292
PerplexityLLM Backbone=Qwen3-8B, Method Category=Training-free methods2026.010.6284
EntropyModel=Llama-2-7B, Generation Phase=Post-generation2026.060.6263
Semantic EntropyLLM Backbone=Llama3.2-3B, Method Category=Training-free methods2026.010.621
MMLLM Backbone=Qwen3-14B, Method Category=Training-based methods2026.010.6191
Lexical SimilarityLLM Backbone=Llama3.2-3B, Method Category=Training-free methods2026.010.6179
PerplexityLLM Backbone=Llama3.1-8B, Method Category=Training-free methods2026.010.6151
Linear probeModel=Llama-2-13B, Generation Phase=Pre-generation2026.060.6104
Linear probeModel=Llama-2-13B, Generation Phase=Post-generation2026.060.6071
MMLLM Backbone=Qwen3-8B, Method Category=Training-based methods2026.010.6044
Attention probeModel=Llama-2-13B, Generation Phase=Post-generation2026.060.6041
Semantic EntropyLLM Backbone=Llama3.1-8B, Method Category=Training-free methods2026.010.6037
SAPLMALLM Backbone=Llama3.1-8B, Method Category=Training-based methods2026.010.6011
Linear probeModel=Mistral-7B, Generation Phase=Pre-generation2026.060.5987
Linear probeModel=Mistral-7B, Generation Phase=Post-generation2026.060.5984
PerplexityLLM Backbone=Llama3.2-3B, Method Category=Training-free methods2026.010.5936
SAPLMALLM Backbone=Qwen3-8B, Method Category=Training-based methods2026.010.5904
SAPLMALLM Backbone=Llama3.2-3B, Method Category=Training-based methods2026.010.5874
MMLLM Backbone=Llama3.1-8B, Method Category=Training-based methods2026.010.5839
A: PerturbationBackbone=Llama-3.3-70B2026.050.57
F: Ans. Expect.Backbone=Llama-3.3-70B2026.050.57
Attention probeModel=Llama-2-7B, Generation Phase=Post-generation2026.060.5637
Attention probeModel=Llama-2-7B, Generation Phase=Pre-generation2026.060.562
MMLLM Backbone=Llama3.2-3B, Method Category=Training-based methods2026.010.5587
SEPLLM Backbone=Llama3.1-8B, Method Category=Training-based methods2026.010.5547
Question lengthModel=Llama-2-7B, Generation Phase=Pre-generation2026.060.5518
ACT (detect)Backbone=Llama-3.3-70B2026.050.55
P(True)Backbone=Llama-3.3-70B2026.050.55
Self-checkModel=Mistral-7B, Generation Phase=Pre-generation, mode=zero-shot2026.060.5485
SelfCheckGPT-NLIBackbone=Llama-3.3-70B2026.050.54
Self-checkModel=Llama-2-7B, Generation Phase=Pre-generation, mode=zero-shot2026.060.5397
Linear probeModel=Llama-2-7B, Generation Phase=Post-generation2026.060.5363
Linear probeModel=Llama-2-7B, Generation Phase=Pre-generation2026.060.5358
Question lengthModel=Llama-2-13B, Generation Phase=Pre-generation2026.060.5301
SEPLLM Backbone=Qwen3-8B, Method Category=Training-based methods2026.010.5253
Question lengthModel=Mistral-7B, Generation Phase=Pre-generation2026.060.5247
SEPLLM Backbone=Qwen3-14B, Method Category=Training-based methods2026.010.5239
EntropyModel=Llama-2-13B, Generation Phase=Post-generation2026.060.5234
Self-checkModel=Llama-2-13B, Generation Phase=Pre-generation, mode=zero-shot2026.060.5233
C: Sem. EntropyBackbone=Llama-3.3-70B2026.050.52
DRIFT-logpBackbone=Llama-3.3-70B2026.050.52
HalluShiftBackbone=Llama-3.3-70B2026.050.52
Log-ProbBackbone=Llama-3.3-70B2026.050.52
MINDBackbone=Llama-3.3-70B2026.050.51
HaloScopeBackbone=Llama-3.3-70B2026.050.51
Ensemble A+C+E+FBackbone=Llama-3.3-70B2026.050.5
INSIDEBackbone=Llama-3.3-70B2026.050.5
IRIS⋆ (self-verify)Backbone=Llama-3.3-70B2026.050.5
DoLaBackbone=Llama-3.3-70B2026.050.5
SEPLLM Backbone=Llama3.2-3B, Method Category=Training-based methods2026.010.4988
SEPsBackbone=Llama-3.3-70B2026.050.49
E: DRIFTBackbone=Llama-3.3-70B2026.050.47
DRIFT-concat (ablation of E)Backbone=Llama-3.3-70B2026.050.47
HaMIBackbone=Llama-3.3-70B2026.050.46
SAPLMABackbone=Llama-3.3-70B2026.050.46
PRISMBackbone=Llama-3.3-70B2026.050.39