Papers
arxiv:2607.04410

AI Wizards at EXIST 2026: Hierarchical Soft-Label Learning for Multimodal Sexism Identification in Memes

Published on Jul 5
· Submitted by
Matteo Fasulo
on Jul 7
Authors:
,
,

Abstract

A multimodal sexism identification system for memes uses hierarchical conditional soft-label prediction with vision-language embeddings and a lightweight Gated MLP trained via KL divergence and uncertainty weighting.

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

Community

Paper author Paper submitter

AI Wizards at EXIST 2026: Hierarchical Soft-Label Learning for Multimodal Sexism Identification in Memes

This paper introduces a hierarchical framework for multimodal sexism detection in memes that explicitly models annotator disagreement instead of reducing subjective annotations to a single label. Built on frozen Gemini Embedding 2 representations, the approach combines a lightweight 3.5M-parameter Gated MLP, KL-divergence training on empirical label distributions, and homoscedastic uncertainty weighting to jointly solve sexism detection, intent classification, and fine-grained sexism categorization.

Rather than fine-tuning large vision-language models, the method pairs frozen multimodal embeddings with hierarchical conditional decoding and loss masking to enforce task dependencies while remaining computationally efficient. On the EXIST 2026 benchmark, it ranked 1st on the fine-grained sexism categorization task and 4th on the other two Soft-Soft leaderboard tasks, highlighting the effectiveness of combining hierarchical learning with disagreement-aware supervision for subjective multimodal content moderation.

Sign up or log in to comment

Get this paper in your agent:

hf papers read 2607.04410
Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2607.04410 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2607.04410 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2607.04410 in a Space README.md to link it from this page.

Collections including this paper 1