A new benchmark called TUPLE-Bench has been developed to evaluate the privacy and utility trade-offs in typed decision models. These models often reveal sensitive information beyond the intended coarse operational category, such as fine-grained causes or probabilities. Experiments across nine models showed that even with explicit instructions to withhold causes, the models frequently leaked this information, with 91.88%-99.93% of responses revealing the cause. While full outputs generally reduce downstream costs, confidence fields increase leakage without consistent cost reduction. The research also introduces workflow-constrained privacy quantization (WCPQ) as a method to mitigate this leakage while preserving downstream workflow utility. AI
IMPACT Highlights the need for careful auditing of AI decision interfaces to prevent unintended information leakage.
RANK_REASON The cluster contains an academic paper detailing a new benchmark and research findings. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
