interpretabilitycritiquebearishThere is a lack of controllable and attributable methods for analyzing how language models resolve conflicts between competing specificationsArtificial Intelligence02 Aug 2026http://arxiv.org/abs/2607.28384v1