Research map / Large language models / Language, safety and interpretability

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

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 SelfCI, a complementary self-distillation method that uses two feedback-conditioned self-teachers (allow and disallow) and reverse KL to reach a product-of-experts target for contextual integrity alignment.

Problem

LLM personal assistants over-disclose or over-suppress private information; aligning them to contextual integrity (CI) requires jointly retaining task-necessary information and suppressing inappropriate information, without degrading utility.

Approach

The model generates its own norm-grounded rationales for allowed and disallowed attributes, conditions itself on each feedback set to form two teachers, and minimizes a weighted sum of reverse KLs to both, which is equivalent to matching a product-of-experts target on the intersection of task completeness and minimal disclosure.

Key results

• On Qwen2.5-7B-Instruct, SelfCI raises Integrity from 35.34 to 83.56 and Complete from 23.29 to 53.42 on CI-RL test.
• On Qwen3-4B (reasoning), Complete improves from 26.03 to 57.26 and Integrity from 32.88 to 82.19 on CI-RL test.
• On PrivacyLens with Qwen3-4B-Instruct, LR drops from 56.59 to 47.06 while Helpful rises to 2.62.

Surprising findings

• Naive self-imitation on the model's own teacher responses (offline self-distillation) degrades utility, indicating overfitting to self-generated responses.
• CI-RL gains do not persist at 4B and 8B scales, remaining near the initial model.

Limitations

• Relies on structured synthetic data with explicit attribute annotations that may not capture real-world CI ambiguity.
• Depends on the model's in-context learning ability, limiting gains for very small models such as Qwen3-0.6B.
• Uses a static lambda; adaptive coefficient selection left to future work; evaluation focuses on final responses, not reasoning traces.

Experimental setup

Datasets: CI-RL PrivacyLens CIMemories

Benchmarks: CI-RL PrivacyLens CIMemories

Metrics: Integrity Utility Complete LR ALR Helpful Violation@5

Baselines: Initial CI-RL ContextDistill

Builds on: self-distillation context distillation GRPO CI-RL product of experts

Closest papers by meaning

• Contextual Integrity in LLMs via Reasoning and Reinforcement Learning (NeurIPS 2025 · similarity 0.93)
• Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory (ICLR 2024 · similarity 0.90)
• PrivAct: Internalizing Contextual Privacy Preservation via Multi-Agent Preference Training (ICML 2026 · similarity 0.89)
• Position: Contextual Integrity is Inadequately Applied to Language Models (ICML 2025 · similarity 0.89)
• Contextualized Privacy Defense for LLM Agents (ICML 2026 · similarity 0.89)
• From History to State: Constant-Context Skill Learning for LLM Agents (NeurIPS 2026 · similarity 0.87)