Research map / Large language models / Reasoning and chain of thought
PrivGate: Steering Contextual Integrity in LLMs via Latent Space Geometry
Runshan Hu, Yukun Dong, Yingying Huangfu, Ruohan Zhao, Yi Xie, Tieyan Li
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PrivGate, which selectively triggers explicit CI-constrained reasoning in LLM agents based on internal latent-space privacy signals, balancing latency and privacy.
Problem
LLM agents must respect contextual integrity (CI) for privacy, but direct generation is fast and privacy-unsafe while CI-constrained explicit thinking is slow; existing agents struggle to balance the two.
Approach
Gate explicit reasoning on-demand using privacy signals read from the model's internal latent space geometry, rather than always reasoning or never reasoning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| gate CI reasoning with latent privacy probe instead of always using direct generation | PrivGate · PrivacyLens end-to-end | leakage rate (%): 70.4 -> 26.3 | Improved |
| gate reasoning with latent probe instead of always-on full CoT | PrivGate · PrivacyLens end-to-end | output tokens: 390 -> 149 | Improved |
| use latent probe gating instead of keyword gating (ablation) | PrivGate · PrivacyLens end-to-end | leakage rate (%): 53.2 -> 26.3 | Improved |
| add gating on top of direct generation | PrivGate · PrivacyLens end-to-end | utility score (%): 97.0 -> 95.0 | Hurt |
| replace entropy-based routing signal with linear latent probe (ablation) | PrivGate · PrivacyLens OOD risk identification | OOD AUROC: 0.78-0.94 -> >0.99 | Improved |
| replace linear probe with 2-layer MLP probe (ablation) | PrivGate probe · PrivacyLens OOD risk identification | OOD AUROC | No clear effect |
| train probe on structure-decoupled data (hard positives/negatives) instead of easy data only (ablation) | PrivGate probe · semantic and template shift tests | OOD AUROC | Improved |
| increase training samples of probe to 64 (ablation) | PrivGate probe · PrivacyLens OOD | generalization accuracy: > 99% with 64 samples | Improved |
| probe intermediate layers (40-70% depth) rather than first/last layers (ablation) | PrivGate probe · PrivacyLens OOD | OOD AUROC | Improved |
Closest papers by meaning
- PrivAct: Internalizing Contextual Privacy Preservation via Multi-Agent Preference Training (ICML 2026 · similarity 0.92)
- Contextual Integrity in LLMs via Reasoning and Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory (ICLR 2024 · similarity 0.90)
- It Takes Two: Complementary Self-Distillation for Contextual Integrity in LLMs (NeurIPS 2026 · similarity 0.90)
- MINIM: Privacy-Aware Minimal View for Agents via Trusted Local Sanitization (ICML 2026 · similarity 0.88)
- POLAR-Bench: A Diagnostic Benchmark for Privacy-Utility Trade-offs in LLM Agents (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
