interpretabilityfactbullishA symmetry-based experimental framework can systematically compare model preferences across representation types by reducing confounding factorsArtificial Intelligence02 Aug 2026http://arxiv.org/abs/2607.28384v1