Logit Bias Constraint

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

• Bias Amplificationconcept0.736
  Problem cited as a limitation of current LLMs; PRH predicts larger models should amplify bias less
• Bias in language modelsconcept0.734
  Features related to gender, racial, ethnic biases, slurs, and hate speech.
• Fluency Constraintconcept0.734
  Early stopping in alpha sweeps when mean fluency falls below 95% of baseline or >5% of responses fall below 90% baseline
• Inductive Biasconcept0.730
  Assumptions or preferences (e.g., parsimony) that determine how a learning system generalizes beyond training data
• Logistic regression correctness probemethod0.730
  Logistic regression trained on GSM8k training set to predict answer correctness from projection features along reflection direction
• Logit-based self-reportmethod0.727
  Primary self-report measure: probability-weighted expected value over all ten digit-token logits, yielding a continuous rating that preserves full distributional signal
• Logit Weight Analysismethod0.725
  Computing each feature's linear effect on output token logits via path expansion through MLP output weights and unembedding matrix
• Binary Generation Constraintconcept0.724
  Implementation technique zeroing all logits except Yes/No tokens to convert steering into binary cross-entropy