Hard Prompt Compression Causes Referential Dangling in Multi-Hop QA

Hard prompt compression techniques frequently suffer from referential dangling, where independent token selection retains an answer but deletes its defining entity. Testing on Qwen3-0.6B and HotpotQA shows error rates as high as 60% for bridge-style questions.