interpretabilityfactneutralPersonality information in LLMs is concentrated in middle transformer layersComputation and Language28 Jul 2026http://arxiv.org/abs/2607.20803v1