Arabic feature A/1/3450 and B/1/1334 have activation correlation of 0.91 across 40M tokens

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Hebrew feature A/1/416 and B/1/1901 have activation correlation of 0.92finding0.863
  Universality of Hebrew script feature across two transformers
• Base64 feature A/1/2357 and B/1/2165 have activation correlation of 0.85finding0.854
  Universality of base64 feature across two transformers
• DNA feature A/1/2937 and B/1/3680 have activation correlation of 0.92finding0.847
  Universality of DNA feature across two transformer models with different random seeds
• Pearson correlation of 0.74 between A/1/3450 activation and Arabic script proxy over 40M tokensfinding0.837
  Joint measure of sensitivity and specificity for the Arabic script feature
• Arabic script feature A/1/3450 fires on 81% Arabic-script tokens when active, with 98% specificity at high activation levelsfinding0.834
  Demonstrates activation specificity of the Arabic script sparse autoencoder feature
• Arabic feature A/1/3450 has 27 neurons with coefficient magnitude ≥0.1 and three largest coefficients are negative; most correlated neuron responds to mixture of non-English languagesfinding0.819
  Demonstrates that the Arabic feature is not aligned to any single neuron
• Feature pair A/1/3949 and B/1/3321 have activation correlation 0.98 but negative logit weight correlation, firing on PLOSOne journal abbreviationsfinding0.811
  Demonstrates that activation similarity can diverge from logit weight similarity due to interference
• Features in A/1 have median activation correlation of 0.72 with most similar feature in B/1; neurons have median 0.46finding0.803
  Systematic comparison showing features are substantially more universal than neurons across models