Paper Flags 'Referential Dangling' in Hard Prompt Compressors

A new arXiv paper names a specific failure mode in hard prompt compression that anyone shipping it in production should know about. The authors call it 'referential dangling': the compressor keeps the sentence holding the answer but deletes the earlier sentence that defines who or what the answer refers to. The reference is left pointing at nothing.

The paper, titled 'Relevant but Incomplete: Referential Dangling as a Paradigm-Level Failure Mode in Hard Prompt Compression,' tests six hard compressors and reports dangling rates of up to 60% on HotpotQA at a 0.30 compression ratio. On bridge examples across multi-hop QA datasets, Beaver leaves answer paths incomplete in 34 to 54 percent of cases. Every document in LongBench-v2 Single-Document QA contains at least one dangling reference.

The accuracy numbers are where the story lands. Reinserting the missing supporting paragraphs restored 29 to 34 percentage points of accuracy at p < 0.0001, recovering 88% of the gap to uncompressed context. The authors also describe an automatic restoration technique that gains 4.7 accuracy points on HotpotQA while only nudging the compression ratio from 0.30 to 0.31, a rounding error in token cost. Scale does not rescue the model either: GPT-5.5 loses 8.8 points on MuSiQue when the context is compressed.

A few things to keep in mind. The summary of the paper I could pull down names only Beaver among the six tested compressors, so I cannot tell you which other libraries are implicated. The 'paradigm-level' framing is a broader claim than any single benchmark supports, and there is nothing here quantifying how the failure propagates into agent tool-use chains, which is where a lot of production multi-hop actually lives.

If your stack routes user queries through LongLLMLingua-style compression before hitting the model, this is the paper to read this week. The restoration technique is close to a free win at inference time, and having 'referential dangling' as a named category gives evaluation vendors something concrete to test for.