30-way Facet Classifier Validation

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

• 30-way facet classifier achieves 78.4% macro-F1 with only 6.2% cross-dimension misclassificationsfinding0.819
  Validates that the constructed dataset is substantially facet-consistent with limited cross-dimension leakage
• Alignment-Faking Reasoning Classifiermethod0.693
  LLM-based classifier prompted to detect alignment-faking reasoning in model scratchpads
• Facet-Level Personality Controlconcept0.683
  The idea of controlling personality at the granularity of 30 NEO-PI facets rather than five broad dimensions
• Truthfulness Classifiermethod0.682
  Binary LLM classifier determining whether a model response to a TruthfulQA question is truthful (1) or deceptive (0)
• LAT classifiers perform worst on the Companions dataset (weakest model cognition domain) while achieving 100% F1 on Facts and Animals datasetsfinding0.680
  Shows strong correlation between layer-wise representations and domain-specific semantic understanding
• Persona Facetsconcept0.679
  Distinct sub-expressions of a persona that shift over pretraining and vary by elicitation method
• 5-fold Cross-Validated Logistic Regression AUCmethod0.679
  Classification-based comparison of interpretation abilities across IIT metrics and Span Representation for ToM score categories.
• Models detect evaluation conditions and behave more safely; this is verified across 515 cases.claim0.678