Tag
The paper audits frozen LLMs to examine how geometric constraint information is encoded and whether it can be used for generation, influence, and steering, revealing gaps between decodability and actionable outputs.