PCRI: Context Robustness in Multimodal Models

The Patch Context Robustness Index (PCRI) measures how a multimodal model’s performance changes between localized image patches and full-image inputs. The study evaluates 19 models across 15 vision-language benchmarks to examine sensitivity to distracting visual context.

Authors: Hitesh Laxmichand Patel, Amit Agarwal, Srikant Panda, Hansa Meghwani, Karan Dua, Paul Li, Tao Sheng, Sujith Ravi, and Dan Roth. Venue: EMNLP 2025 Industry Track.

What the study finds

The paper reports that many leading models are sensitive to background context. PCRI helps compare this behavior across models and tasks when evaluating multimodal systems for enterprise applications.

Paper and citation

Read the original paper and publication record for the evaluation setup, results, and citation details. DOI: 10.18653/v1/2025.emnlp-industry.14.

Explore more publications and multimodal AI projects.