Relevant but Incomplete: Referential Dangling as a Paradigm-Level Failure Mode in Hard Prompt Compression

Part of

Vision-Language-Action And Multimodal Models