Arabic script feature A/1/3450 fires on 81% Arabic-script tokens when active, with 98% specificity at high activation levels

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

• Arabic feature A/1/3450 and B/1/1334 have activation correlation of 0.91 across 40M tokensfinding0.834
  Demonstrates universality of the Arabic script feature across two independently trained transformers
• Pearson correlation of 0.74 between A/1/3450 activation and Arabic script proxy over 40M tokensfinding0.800
  Joint measure of sensitivity and specificity for the Arabic script feature
• The Arabic script feature would be effectively invisible if we only analyzed the model in terms of neurons.quote0.752
  Summarizes key finding that monosemantic features cannot be discovered by neuron-level analysis
• Hebrew feature A/1/416 and B/1/1901 have activation correlation of 0.92finding0.744
  Universality of Hebrew script feature across two transformers
• In A/4, over 100 features primarily respond to the token 'the' in different contextsfinding0.740
  Demonstrates prevalence of token-in-context features and feature splitting of common tokens
• 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 languagesfinding0.731
  Demonstrates that the Arabic feature is not aligned to any single neuron
• Pinning A/1/3450 to maximum observed value causes model to generate Arabic text from numeric prefix contextfinding0.728
  Causal validation that the Arabic feature has the predicted downstream effect on generation
• Feature 1M/1013764 activates on diverse code errors (typos in code, array overflow, divide by zero, type mismatch) across Python, C, Scheme, but not on English prose typos.finding0.728
  Shows a general code error detector beyond simple typo detection.