TUPLE-Bench: Evaluating and Mitigating Privacy-Utility Trade-offs in Typed Decision Models
We introduce TUPLE-Bench (Typed decision models' Utility and Privacy Leakage Evaluation Benchmark), a benchmark of paired synthetic cases in which recipients may learn a coarse operational category but not its fine-grained cause.
ProofPaper ↗
Key points
- Typed decision models often return more than a selected category.
- These fields support downstream decisions but may reveal information beyond a recipient's authorization.
- We further introduce workflow-constrained privacy quantization (WCPQ), which reduces secret recovery while preserving the behavior and cost of specified downstream workflows.
- Our findings highlight the need to audit decision interfaces field by field and workflow by workflow.
Sources (1)
- [1]TUPLE-Bench: Evaluating and Mitigating Privacy-Utility Trade-offs in Typed Decision ModelsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 04:00 PM
We introduce TUPLE-Bench (Typed decision models' Utility and Privacy Leakage Evaluation Benchmark), a benchmark of paired synthetic cases in which recipients may learn a coarse operational category but not its fine-grained cause.
Typed decision models often return more than a selected category.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 16, 2026Chinese Decision: A Family of Language Models for Single-Pass Probabilistic Decision-Making
- Oct 16, 2026Welcome Beam: a 501B open model you can own, customize, and scale
- Oct 15, 2026NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
- Oct 15, 2026Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio
- Oct 15, 2026Welcome FLUX.3 Image Klein in Diffusers
- Oct 14, 2026Grok Voice 3.0 from SpaceXSI outscores OpenAI and Google on speech-to-speech index