MMS-e: Benchmarking the Resilience of Large Multimodal Models to Visual Scrambling

Benchmark Examples

Patchwise Question Answering: Divide the images into 2x2, 4x4, and 8x8 patches, then shuffle all the patches, and measure the ability of LMMs to answer questions about these images. Fixed Patch Question Answering: Divide the image into 4x4 patches, randomly fix some of the patches, and let LMMs answer questions based on the image. Reconstruction task: Let LMMs reconstruct the order of shuffled patches based on the image' s caption, and let LMMs reconstruct the shuffled caption based on the image.

Directory Structure

• Patchwise QA/:
  • The images about Patchwise Question Answering.
• Reconstruction/:
  • The images about Reconstruction task.
• Fixed Patch QA/:
  • The images about Fixed Patch Question Answering.