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

2026-05-22 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Robotics & Autonomous Systems · Depth: Expert, long

Summary

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 (LLMs) 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, SelfCI is proposed as 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) on both in-domain (CI-RL test set) and out-of-domain (PrivacyLens benchmark) settings.

Key takeaway

For Machine Learning Engineers developing LLMs for sensitive personal agent workflows, SelfCI offers a practical solution to the persistent privacy-utility trade-off. You should consider implementing this complementary self-distillation framework to ensure contextual integrity without sacrificing task performance. By decoupling information suppression from task resolution through distinct teacher distributions, your models can achieve both task completeness and minimal disclosure. This approach avoids costly external supervision and outperforms traditional online reinforcement learning, providing a robust path toward CI alignment in real-world applications.

Key insights

SelfCI decouples privacy and utility in LLMs using complementary self-distillation to achieve Contextual Integrity, outperforming RL baselines.

Principles

• Privacy is context-dependent information flow, not mere secrecy.
• Decouple information suppression from task resolution for CI.
• Jointly optimize distinct objectives to resolve privacy-utility trade-offs.

Method

SelfCI generates rationales to condition two teacher policies (task-completion, minimal disclosure) and jointly optimizes a student against their Product-of-Experts target using reverse KL divergences.

In practice

• Generate self-feedback to guide context-specific disclosure decisions.
• Apply Product-of-Experts to balance task utility and privacy constraints.

Topics

• Contextual Integrity
• Large Language Models
• Self-Distillation
• Privacy-Preserving AI
• Product-of-Experts
• Information Flow Governance

Code references

• sw-programmer/SelfCI

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Microsoft Research Develops Novel Approaches to Enforce Privacy in AI Models — Contextual integrity in LLMs can be enforced via inference-time checks or advanced training methods to prevent sensitive information disclosure.
• SHRED: Retain-Set-Free Unlearning via Self-Distillation with Logit Demotion — LLM unlearning can be retain-set-free by selectively demoting high-information tokens via self-distillation, preserving general utility.
• TriAlign: Towards Universal Truth Consistency in Personalized LLM Alignment — TriAlign ensures personalized LLMs maintain universal truth consistency across diverse social groups while preserving individual preferences.
• Taking LLMs out of the black box: A practical guide to human-in-the-loop distillation — Human-in-the-loop distillation enables deploying LLM knowledge into smaller, transparent, in-house models.
• Reinforcing privacy reasoning in LLMs via normative simulacra from fiction — Fiction-derived normative simulacra and GRPO can teach LLMs transferable contextual privacy reasoning.
• Benchmarking Empirical Privacy Protection for Adaptations of Large Language Models — Distribution shifts critically impact practical privacy in differentially private LLM adaptations, often undermining theoretical guarantees.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.