List your company

Concept Faithfulness Evaluation on Jigsaw

91.72Perturbed AccuracySAE
54.862464.43127483.5688Jun 24, 2025

Evaluation results

MethodLinks
SAEModel=RoBERTa2025.0691.72-0.6
Single-LayerModel=RoBERTa2025.0690.191.1
SAEModel=Qwen2025.0684.630.2
SAEModel=Llama2025.0684.10.3
SAEModel=BERT2025.0683.357.3
Single-LayerModel=BERT2025.0680.7910.2
CLVQ-VAEModel=Llama2025.0679.315.9
ClusteringModel=Llama2025.0677.937.5
Single-LayerModel=Llama2025.0676.229.6
ClusteringModel=BERT2025.0675.7715.8
CLVQ-VAEModel=BERT2025.0673.7218
ClusteringModel=Qwen2025.0663.5223.6
CLVQ-VAEModel=RoBERTa2025.0661.2732.8
Single-LayerModel=Qwen2025.0659.1728.8
CLVQ-VAEModel=Qwen2025.0658.0930.1
ClusteringModel=RoBERTa2025.0656.2838.3