Tsinghua Reveals 'Referential Dangling' Flaw in Hard Prompt Compression
Tsinghua · hf · 2026-08-10
Research from Tsinghua University identifies a structural failure in hard prompt compression called 'referential dangling.' When algorithms independently score and retain high-relevance text chunks, they often split dependent evidence pairs. This leaves answers without the necessary entity definitions, making them uninterpretable.
Experiments show that at a 0.30 compression ratio, all tested mainstream compressors exhibit dangling rates up to 60% on multi-hop QA datasets. Even advanced models like GPT-5.5 cannot absorb this loss, showing an 8.8 percentage point accuracy drop compared to full contexts.
To fix this, the researchers trained a compact classifier to rank omitted sentences and automatically reinsert missing referential backgrounds during inference. This restoration improves accuracy by 4.7 points on HotpotQA while barely changing the compression ratio (from 0.30 to 0.31).
More from Research
- Stanford's ChatEHR Deployment: $6M+ Estimated First-Year ROI and Inadequacies of Benchmarks — EricTopol · 2026-08-10
- ICML 2026 Machine Unlearning Tutorial Released with Videos and Slides — thegautamkamath · 2026-08-10
- SceneGen Generates 3D Scenes from a Single Image in One Feedforward Pass — tom_doerr · 2026-08-10
- Surya Ganguli Shares Top Summer Schools in Computational Neuroscience and AI — SuryaGanguli · 2026-08-10
- IFDS Research Expo Aug 10-12 at UW-Madison: AI foundations, ML, stats, optimization — prof_kamilov · 2026-08-10
- Primus Launches Autonomous ML Research Agent: 30x Faster from Hypothesis to Paper — JayAlammar · 2026-08-10