Typicality bias rates in instruction-tuned models remain at similar or higher levels compared to their base model counterparts

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

• Typicality bias is present across all annotator demographic groups but varies in magnitude, potentially reflecting differences in linguistic norms captured by pretraining dataclaim0.858
  Finding from PRISM dataset analysis showing typicality bias varies by ethnicity and region, with implications for fairness
• Typicality bias rate exceeds 50% chance baseline by 4-12 percentage points across all base models on all four preference datasetsfinding0.841
  Systematic evidence that base models implicitly prefer human-preferred responses, indicating preference biases emerge during pretraining
• Typicality Bias Ratemethod0.833
  Measurement of how often human annotators prefer the response with higher base model log-probability
• Typicality Biasconcept0.803
  The human tendency to prefer more typical, familiar, fluent, and predictable text in annotation tasks, identified as a fundamental data-level cause of mode collapse
• We hypothesize that native self-report, fine-tuned introspection models, and trained activation-to-language systems will show different performance on bias-resistant localization and strength benchmarkshypothesis0.797
  Comparative prediction motivating future work contrasting different approaches to LLM self-knowledge
• Typicality bias with weight α>0 mathematically sharpens the aligned policy via a power transform of the reference distributionclaim0.796
  Theoretical result showing that any positive typicality bias weight γ-sharpens the reference distribution, amplifying modes
• Larger models should amplify bias less than smaller models, with model biases more accurately reflecting data biases rather than exacerbating themclaim0.796
  Implication of PRH for AI fairness and bias
• Models fine-tuned on human data show strong positive correlation between misalignment and incoherence; misalignment scores fall below 1% when incoherence threshold is appliedfinding0.791
  Human data fine-tuning effect is distinct from synthetic emergent misalignment and likely caused by off-policy training