TUPLE-Bench: Evaluating privacy-utility trade-offs in typed decision models
In brief
- Typed decision models leak fine-grained details despite privacy instructions.
- TUPLE-Bench measures sensitive data leaks across 9 models with 91.88% to 99.93% exposure rates.
- WCPQ reduces leaks while preserving workflow efficiency.
A study shows that typed decision models often leak sensitive details despite privacy instructions. The TUPLE-Bench benchmark reveals that 91.88% to 99.93% of responses expose forbidden fine-grained information. A quantization method, WCPQ, reduces these leaks without harming workflow efficiency.
This summary gives you the essentials. The full article, its figures and quotes are on cs.AI updates on arXiv.org.
Read the original articleAI-generated summary of the source, checked by rules. Report an error