Abstract
We present the AI Wizards submission to EXIST 2026 for multimodal sexism identification in memes. The system treats the three hierarchical subtasks 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. The submitted systems ranked first in Task 2.3 and fourth in Tasks 2.1 and 2.2 on the official Soft-Soft leaderboards.
Method
The model treats Gemini Embedding 2 vectors as fixed meme representations indexed by the official EXIST identifiers. A compact Gated MLP predicts sexism identification, source intention, and sexism category with conditional masking.
Soft predictions are decoded through the task hierarchy: downstream intention and category probabilities are conditioned on the predicted probability that a meme is sexist. Hard predictions use a simple 0.5 threshold without task-specific calibration.
Fixed Embeddings
Cached Gemini features encode the meme content before training, keeping the task model compact and reproducible.
Gated MLP
A lightweight classifier predicts the three hierarchical subtasks from the shared meme representation.
Soft Labels
KL divergence targets empirical annotator distributions, with learned uncertainty weights balancing tasks.
Task Setting
The EXIST 2026 meme dataset contains 3,984 training instances and 1,053 test instances across English and Spanish. The task follows the Learning with Disagreement setting, so systems predict annotator distributions instead of only majority labels.
The code expects the official EXIST dataset under the local data/ directory. Gemini embedding caches are distributed through the gated Hugging Face repository AIWizards/EXIST2026-embeddings. The model weights are still pending upload and will be published separately.
Generated submission files are stored in exist2026_aiwizards/. The public repository also keeps a frozen copy of the official-format outputs used for evaluation and reproduction.
Official Results
The system was designed for the Soft-Soft setting and achieved its strongest result on Task 2.3, where the three submitted runs ranked first, second, and third.
| Task | System | Rank | ICM-Soft | ICM-Soft Norm |
|---|---|---|---|---|
| 2.1 Sexism identification | aiwizards_3 | 4 | 0.2323 | 0.5373 |
| 2.2 Source intention | aiwizards_1 | 4 | -0.6720 | 0.4285 |
| 2.3 Sexism categorization | aiwizards_2 | 1 | -2.8881 | 0.3469 |
On the local 10% test split, random-seed variance was most visible on the downstream tasks. Base MLP Run 2 was strongest for Task 2.3 soft-label prediction, while Run 1 was stronger for Task 2.2.
Repository
The implementation includes the PyTorch Lightning training loop, dataset loading, soft-target construction, Gemini and SigLIP model variants, checkpoint prediction, and PyEvALL evaluation helpers.
uv sync
uv run python train.py --model_name gemini --version final --epochs 50 --soft_gating
uv run python predict.py --ckpt_path checkpoints/gemini_final/epoch=XX.ckpt --version gemini_final --output_dir exist2026_aiwizardsLimitations
The repository includes optional demographic and physiological ablations, but the public submission path uses the fixed Gemini embeddings and the hierarchical prediction heads described in the paper.
BibTeX
@misc{fasulo2026aiwizardsexist2026,
title={AI Wizards at EXIST 2026: Hierarchical Soft-Label Learning for Multimodal Sexism Identification in Memes},
author={Matteo Fasulo and Antonio Gravina and Luca Tedeschini and Luca Babboni},
year={2026},
eprint={2607.04410},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2607.04410}
}