knobe-effect

This paper investigates whether the Knobe effect (a moral bias in intentionality judgments) emerges in finetuned LLMs, and uses layer-patching analysis to localize the bias to specific layers, showing that targeted interventions can eliminate the effect without retraining.