List your company
ResearchFormerly WizWand

Uncertainty Estimation on CommonsenseQA

0.74AUROCGeometry-Calibrated Conformal Abstention
0.45920.53210.6050.6779Apr 30, 2026
Updated 6mo ago

Evaluation results

MethodLinks
Geometry-Calibrated Conformal AbstentionBackbone=LLaMA-3.2-3B-Instruct2026.040.740.9
ProposedBackbone Model=LLaMA-3-8B-Instruct2026.040.710.9
PerplexityBackbone=LLaMA-3.2-3B-Instruct2026.040.690.87
SARBackbone=LLaMA-3.2-3B-Instruct2026.040.690.88
FocusBackbone=LLaMA-3.2-3B-Instruct2026.040.690.86
EigenBackbone Model=LLaMA-3-8B-Instruct2026.040.690.88
EigenBackbone=LLaMA-3.2-3B-Instruct2026.040.680.87
PerplexityBackbone Model=LLaMA-3-8B-Instruct2026.040.680.88
SARBackbone Model=LLaMA-3-8B-Instruct2026.040.680.88
SemanticBackbone Model=LLaMA-3-8B-Instruct2026.040.680.88
FocusBackbone Model=LLaMA-3-8B-Instruct2026.040.670.86
SemanticBackbone=LLaMA-3.2-3B-Instruct2026.040.660.87
ATRMDBackbone Model=LLaMA-3-8B-Instruct2026.040.630.85
AttentionBackbone Model=LLaMA-3-8B-Instruct2026.040.540.81
ATRMDBackbone=LLaMA-3.2-3B-Instruct2026.040.530.8
P(True)Backbone Model=LLaMA-3-8B-Instruct2026.040.530.82
AttentionBackbone=LLaMA-3.2-3B-Instruct2026.040.50.77
P(True)Backbone=LLaMA-3.2-3B-Instruct2026.040.470.76