Larger numbers of entailment predictions indicate lack of diversity; larger numbers of contradiction predictions indicate higher diversity

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

• # Entailment is negatively correlated with diversity (ρ=-0.48 decTest, ρ=-0.65 conTest)finding0.802
  Ablation result confirming that entailment predictions indicate lack of diversity
• # Contradiction is positively correlated with diversity (ρ=0.26 decTest, ρ=0.46 conTest)finding0.759
  Ablation result confirming that contradiction predictions indicate diversity
• Bigger models are more likely to converge to a shared representation than smaller modelshypothesis0.759
  Selective pressure toward convergence via model capacity
• The contradiction relation is more useful than the neutral relation for measuring semantic diversityclaim0.758
  Key finding from ablation study on NLI prediction categories
• Universality claims for truth directions are more limited than previously assumed, with significant differences observable for various model layers, task difficulties, task types, and prompt templates.claim0.746
  Overarching conclusion summarizing the paper's contribution relative to prior universality claims.
• Confidence NLI Diversity achieves ρ=0.64 correlation with human diversity judgments on conTestfinding0.745
  Highest human correlation for semantic diversity metric
• Adding a single disambiguating example (12−9=21) aligns divergent M1-M4 interpretations under tested seedsfinding0.744
  E1 finding consistent with threshold-crossing: near-threshold state resolved by one additional anchor
• Neutral NLI predictions may capture lexical rather than semantic diversityhypothesis0.744
  Hypothesis proposed to explain Neutral NLI Diversity's high performance on decTest but low on conTest