ResearchBenchmarksInterpretability Evaluation on Top-activated TextsFollow0.74Embedding SimilarityCLIP+SAE-0.10240.11630.3350.5537Feb 16, 2025Evaluation ResultsMethodMethodLinksEmbedding SimilarityWin RateCLIP+SAE2025.020.7461CLIP2025.020.4559CLIP+NCL2025.020.4560DeCLIP2025.02-0.0746