ConCI: A Controlled Counterfactual Benchmark for Contextual Integrity in LLMs

Key points

• However, existing CI benchmarks often use independent scenarios that vary multiple CI parameters simultaneously and are dominated by inappropriate flows, making it unclear whether correct predictions reflect sensitivity to context or a general refusal tendency.1
• We present ConCI, a controlled-counterfactual benchmark in which an anchor flow is paired with two counterfactuals that each modify a single CI parameter.1
• Across seven LLMs, many correct anchor predictions fail on at least one counterfactual, and even the strongest model errs on about a fifth of triplets.1
• Larger models narrow this bias toward non-disclosure but do not eliminate it, and reasoning yields only modest gains.1

Sources (1)

• [1]ConCI: A Controlled Counterfactual Benchmark for Contextual Integrity in LLMs
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 06:29 AM

  Contextual integrity (CI) holds that the appropriateness of sharing personal information depends on the context of the information flow, not on the information alone.

  However, existing CI benchmarks often use independent scenarios that vary multiple CI parameters simultaneously and are dominated by inappropriate flows, making it unclear whether correct predictions reflect sensitivity to context or a general refusal tendency.

Extractive summary: sentences quoted from the sources.