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

Researchers introduced SELFCI, a new self-distillation framework for large language models that separates information suppression from task solving.

The method uses complementary objectives instead of external supervision to better follow contextual integrity rules and control what models disclose.

Reported results show SELFCI outperforms reinforcement-learning baselines on both privacy and utility, including harder out-of-domain tests.

The work addresses a key deployment challenge for AI assistants: protecting sensitive information without reducing usefulness on real tasks.