Skip to content
InfoResearchPreprintLLM-specific

PriTabICL: Differentially Private Tabular In-Context Learning with Task-Aligned Cluster-Wise Synthesis

Published
Record updated
View JSON

Summary

PriTabICL is a framework for generating differentially private demonstrations for tabular in-context learning with LLMs. It partitions records into DP clusters, synthesizes each cluster independently using a fixed feature-target workload, and routes each query to a cluster to sample demonstrations. On four real-world datasets, PriTabICL-PrivSyn beats global PrivSyn in 8 of 12 settings and outperforms existing private TabICL approaches in nearly all evaluated settings.