Ely Hahami, Ishaan Sinha, Lavik Jain, Josh Kaplan, Jon Hahami · 2025 · arXiv preprint Read the source Yes/no injection tests carry a confound — steering pushes models toward ‘yes’. On confound-free tasks, models could localize injected perturbations, but mainly for early layers. Questions it bears on How well can LLMs introspect