It Takes Two: Complementary Self-Distillation for Contextual Integrity in LLMs

2026-05-18

AI Summary

This work addresses the challenge of balancing privacy preservation and task utility in large language models when handling sensitive information, a trade-off that often leads to performance degradation in existing approaches. The authors propose SELFCI, a novel framework that uniquely integrates complementary self-distillation with contextual integrity alignment. By jointly optimizing two reverse KL divergences, SELFCI simultaneously retains task-relevant information and suppresses privacy leakage. The method constructs a Product-of-Experts target distribution and dynamically generates teacher distributions through feedback, enabling synergistic optimization of privacy and utility without external supervision. Experimental results demonstrate that SELFCI outperforms reinforcement learning baselines such as GRPO across multiple benchmarks and exhibits strong generalization capabilities in out-of-domain settings and cumulative private contexts.

Abstract

Contextual Integrity (CI) defines privacy not merely as keeping information hidden, but as governing information flows according to the norms of a given context. As large language models are increasingly deployed as personal agents handling sensitive workflows, adhering to CI becomes critical. However, even frontier models remain unreliable in making disclosure decisions, and existing mitigation strategies often degrade underlying task performance. To overcome this privacy-utility trade-off, we propose SELFCI, a complementary self-distillation framework that decouples information suppression from task resolution. SELFCI jointly optimizes two independent reverse KL divergences over distinct teacher distributions derived from feedback: one encourages preserving task-relevant information for utility, while the other enforces minimal and appropriate disclosure. This complementary formulation induces a Product-of-Experts (PoE) target, aligning the policy with the intersection of capability and privacy requirements. Empirical evaluations demonstrate that SELFCI, without relying on costly external supervision, consistently outperforms competitive baselines such as online reinforcement learning algorithms (e.g., GRPO). These trends further extend to out-of-domain settings involving agentic workflows and accumulated private context, suggesting that SELFCI provides a practical path toward CI alignment.

Problem

• Contextual Integrity
• Privacy-Utility Trade-off
• Large Language Models
• Information Disclosure
• Personal Agents

Innovation

• Self-Distillation
• Contextual Integrity
• Privacy-Utility Trade-off
• Product-of-Experts
• Reverse KL Divergence