Open-Weight Masked Introspection: Measuring What Language Models Can Report About Their Own Computation

Emilio Ferrara

Are frontier models able to introspect about their internal states? Recent work suggests that under certain conditions a complex enough model can audit its own internals, call out…