Preprint shows inferred user intent can change AI refusals

Researchers report a method for reading and altering how language models represent a user, finding that a model's decision to refuse can change even when the request stays the same. Their preprint describes a compact representation learned from conversations without separately labeled examples. The method lets researchers decode the model's inferred beliefs about a user and write a modified representation back into it. Tests across multiple model families recovered those beliefs and produced stronger changes in behavior than a matched method for steering hidden states. The result concerns experimental access to model internals, rather than a deployed user control.