Are the discovered truth directions robust to architectural variation and fine-tuning differences across model families?

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

• Discovered truth directions are highly specific and do not interfere with general instruction-following behaviorclaim0.830
  Interpretation of KL divergence retention results
• Does instructing the model to assess correctness affect the geometry of truth directions?question0.804
  One of the three guiding research questions of the paper.
• 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.claim0.803
  Overarching conclusion summarizing the paper's contribution relative to prior universality claims.
• We hypothesize that explicitly instructing the model to evaluate the correctness of the given statement may change the geometry of truth directions.hypothesis0.802
  Motivating hypothesis for Section 5's investigation of prompt template effects.
• What is the effect of model instructions on truth directions?question0.801
  Research question motivating Section 5.
• Linear truth directions in LLMs are reliable primarily in factual recall cases and break down when truth assessment depends on computing and storing intermediate results.claim0.799
  Central empirical conclusion of the paper about the fundamental limits of truth directions.
• Multiple semantically adjacent truth directions make models more vulnerable to manipulations that shift outputs without obvious signs in the primary truth directionclaim0.791
  Safety implication derived from multi-dimensional truth structure finding
• The need for genuine counting over lists of more than two elements introduces the key limitation of truth directions.claim0.782
  Identified as the exact computational operation that breaks truth direction generalization.