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

Jul 16, 2026

Upload by Violet

Presents the AI Wizards submission for EXIST 2026 focused on multimodal sexism identification in memes. The task is organized as three increasingly difficult subtasks and is modeled hierarchically through conditional soft-label prediction over annotator distributions. Fixed Gemini Embedding 2 vision-language representations are mapped via a lightweight Gated MLP trained with KL divergence and homoscedastic uncertainty weighting. The system achieved first place on Task 2.3 and fourth on Tasks 2.1 and 2.2.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

Introduction

Related Work

Multimodal Sexism Detection

Architectures and Representation

AI Wizards at EXIST 2026: Hierarchical Soft-Label Learning for Multimodal Sexism Identification in Memes Notebook for the EXIST Lab at CLEF 2026 Matteo Fasulo1, *,†, Antonio Gravina2,†, Luca Tedeschini3,† and Luca Babboni4,† 1 Swiss Data Science Center, ETHZürich, Andreasturm, Andreasstrasse 5, 8092 Zürich, Switzerland 2 Everest Systems GmbH, Max-Jarecki-Straße 21, 69115 Heidelberg, Germany 3 Villanova. ai S.P. A, Località Sa Illetta, SS 195 KM 2 . 3, 09123 Cagliari, Italy 4Independent researcher 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](https://github.com/NLP-AI-Wizards/EXIST-2026) . Keywords sexism identification, multimodal memes, learning with disagreement, hierarchical classification, soft labels 1. Introduction Sexist content online reinforces women’s marginalization and exclusion from digital spaces [1] . Digital platforms often amplify offline patriarchal structures, where hostile and benevolent sexism jointly legitimize gender inequality [2] . Exposure to such content causes measurable psychological harm (e.g., elevated anxiety, depression, and self-censorship), which silences women’s participation in public discourse [3] . Robust automated detection is therefore a critical intervention for digital safety. Memes pose unique challenges: their meaning emerges non-linearly from text-image interaction, often relying on irony, humor, or implicit cultural knowledge to provide plausible deniability [4, 5] . Detection systems must model how modalities jointly construct gendered meanings that may be implicit, ironic, or ambiguous. EXIST 2026 [6, 7] structures sexism detection as three hierarchically nested subtasks, detailed in Section 3: binary sexism identification, source intention detection for sexist content, and fine-grained multi-label categorization across five sexism types. Critically, the task adopts the Learning with Disagreement (LeWiDi [8]) paradigm, requiring systems to predict the full empirical annotator distribution rather than a single majority label. We address all three subtasks as hierarchical conditional multi-task learning over pre-computed vision-language embeddings, preserving semantic dependencies while directly optimizing for soft-label evaluation. Our main contributions are as follows. CLEF 2026 Working Notes, 21 – 24 September 2026, Jena, Germany * Corresponding author. † These authors contributed equally. $ [matteo.fasulo@sdsc.ethz.ch](matteo.fasulo@sdsc.ethz.ch) (M. Fasulo); [antonio.gravina@everest-erp.com](antonio.gravina@everest-erp.com) (A. Gravina); [ltedeschini@villanova.ai](ltedeschini@villanova.ai) (L. Tedeschini); [luca.babboni2@studio.unibo.it](luca.babboni2@studio.unibo.it) (L. Babboni) 􀂀 https://matteofasulo.com (M. Fasulo); [https://github.com/GravAnt](https://github.com/GravAnt) (A. Gravina); [https://github.com/LucaTedeschini](https://github.com/LucaTedeschini)[ ](https://github.com/LucaTedeschini)(L. Tedeschini); [https://github.com/ElektroDuck](https://github.com/ElektroDuck) (L. Babboni) 􀀚 0000-0002-7019-3157 (M. Fasulo); 0009-0007-9252-263X (A. Gravina); 0009-0006-0375-829X (L. Tedeschini); 0009-0001-5260-7467 (L. Babboni) © 2026 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0) . • Multimodal representation: Pre-trained Gemini Embedding 2 features processed through compac

FAQs

What is the overall goal of the AI Wizards submission for EXIST 2026?

To identify sexist content in memes using multimodal information from text and images, structured as three hierarchically nested subtasks.

How does the system handle the requirement to predict annotator disagreement?

It predicts the full empirical annotator distribution using hierarchical conditional soft-label learning rather than a single majority label.

What model and training approach are used to produce soft-label predictions?

It maps fixed Gemini Embedding 2 representations through a lightweight Gated MLP, trained with KL divergence and homoscedastic uncertainty weighting.

Jul 16, 2026

Upload by Violet