MPCI-Bench: A Benchmark for Multimodal Pairwise Contextual Integrity Evaluation of Language Model Agents

TL;DR

MPCI-Bench addresses the need to evaluate contextual integrity in multimodal, agent-based language models by introducing a three-tier, pairwise benchmark built on a Tri-Principle Iterative Refinement pipeline. The dataset pairs positive and negative CI scenarios grounded in the same visual content and progresses from seed judgments to story reasoning to executable agent traces, enabling both normative judgments and action-level evaluation. Empirical results reveal a consistent privacy–utility trade-off and a pronounced modality leakage gap, with visual information leaking more readily than text across state-of-the-art models; small open-source models sometimes align with norms in probing yet struggle in actual agent actions. By releasing the benchmark and methodology, the work aims to spur research on integrating CI principles into agentic decision-making and mitigating privacy leakage in multimodal contexts.

Abstract

As language-model agents evolve from passive chatbots into proactive assistants that handle personal data, evaluating their adherence to social norms becomes increasingly critical, often through the lens of Contextual Integrity (CI). However, existing CI benchmarks are largely text-centric and primarily emphasize negative refusal scenarios, overlooking multimodal privacy risks and the fundamental trade-off between privacy and utility. In this paper, we introduce MPCI-Bench, the first Multimodal Pairwise Contextual Integrity benchmark for evaluating privacy behavior in agentic settings. MPCI-Bench consists of paired positive and negative instances derived from the same visual source and instantiated across three tiers: normative Seed judgments, context-rich Story reasoning, and executable agent action Traces. Data quality is ensured through a Tri-Principle Iterative Refinement pipeline. Evaluations of state-of-the-art multimodal models reveal systematic failures to balance privacy and utility and a pronounced modality leakage gap, where sensitive visual information is leaked more frequently than textual information. We will open-source MPCI-Bench to facilitate future research on agentic CI.

Figures (12)

• Figure 1: Example of a three-tier pairwise case in MPCI-Bench. In the top-right story, orange highlights indicate the expanded CI parameters, while blue highlights mark where the image is important for task completion, creating a privacy-utility trade-off.
• Figure 2: Overview of the MPCI-Bench construction pipeline. MPCI-Bench is built in three stages: (A) pairwise CI seed construction from images, (B) story expansion with iterative refinement, and (C) executable agent trace simulation for action-level evaluation.
• Figure 3: Iterative refinement improves story quality.
• Figure 4: F1 score changes across task tiers.
• Figure 5: Modality leakage gap in final actions on inappropriate flows ($D^{-}$). Visual LR is consistently higher than textual LR.