Benchmark RadarPragMatch
PragMatch is a controlled set of 3,000 image-text pairs derived from MMSD2.0 for studying pragmatic incongruity in multimodal sarcasm detection, with original sarcastic examples and constructed literal and hard-negative pairs.
- Released
- 2026-08-10
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
This resource helps investigate whether large vision-language models rely on superficial cues instead of genuine reasoning in multimodal sarcasm, highlighting practical limitations in model evaluation.
Motivation
Large Vision-Language Models (LVLMs) have demonstrated strong performance on multimodal benchmarks, yet it remains unclear whether they genuinely reason about relationships between images and text or rely on superficial correlations, known as shortcut learning.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.