Megadose The AI curve, compressed.

Abstention Direction in Large Language Models: Mechanistic Interpretability of Abstention Across Query Ambiguity, Knowledge Limits, and Safety Alignment

· HF Daily Papers ·
Steering internal model activations changed abstention across ambiguous questions, knowledge gaps and safety concerns, with effects carrying across categories.

The researchers built matched query pairs that differed only in the factor expected to make a model withhold an answer. They used differences in mean activations to identify an abstention direction for each category. Interventions in middle layers increased or decreased abstention across all three categories. The cross-category effects indicate that these directions share a common directional component.

HF Daily Papers' note

score 4

Categories: Research