🤖 AI Summary
This work addresses the challenge of multimodal gender bias detection in internet memes by proposing a hierarchical soft-label learning framework that models three progressively difficult subtasks as conditional prediction problems, effectively capturing both inter-annotator semantic disagreements and task dependencies. Leveraging visual-language representations from Gemini Embedding 2, the approach introduces a lightweight gated MLP and employs an optimization strategy combining KL divergence with homoscedastic uncertainty weighting. Evaluated on the EXIST 2026 benchmark, the method achieves first place on Task 2.3 and fourth place on both Tasks 2.1 and 2.2 in the official Soft-Soft leaderboard, demonstrating its effectiveness and technical novelty.
📝 Abstract
We present the AI Wizards submission to EXIST 2026 for multimodal sexism identification in memes. The task is composed of three, increasingly harder subtasks. We model them hierarchically as conditional soft-label prediction over empirical annotator distributions. Our system maps fixed Gemini Embedding 2 vision-language representations through a lightweight Gated MLP trained with KL divergence and homoscedastic uncertainty weighting. Our submissions ranked first on Task 2.3 and fourth on Tasks 2.1 and 2.2 on the official Soft-Soft leaderboards. The code is available at https://github.com/NLP-AI-Wizards/EXIST-2026