- The paper presents a 3.5M-parameter hierarchical model that combines frozen Gemini Embedding 2 representations, KL-divergence soft-label training, uncertainty-based task weighting, and conditional decoding.
- The system ranked first on EXIST 2026’s fine-grained multi-label sexism task, fourth on binary sexism and intention detection, and achieved a Soft-Soft ICM score of −2.8881 on Task 2.3.
- The results show that soft-label prediction suits annotator disagreement better than untuned hard-label thresholds, while EEG signals provided negligible linear separability and were excluded from the final model.
This paper describes the AI Wizards submission to the EXIST 2026 shared task on multimodal sexism identification in memes, organized as a CLEF 2026 lab. The system addresses three hierarchically nested subtasks—binary sexism identification (Task 2.1), source intention detection conditioned on sexist content (Task 2.2), and fine-grained multi-label sexism categorization across five categories (Task 2.3)—under the Learning with Disagreement (LeWiDi) paradigm, which requires predicting the full empirical annotator distribution rather than a single majority label. The approach combines frozen Gemini Embedding 2 vision-language representations, a lightweight Gated MLP backbone, KL-divergence training against soft targets, and homoscedastic uncertainty weighting for dynamic task balancing. On the official Soft-Soft leaderboards, the system ranked first on Task 2.3 and fourth on Tasks 2.1 and 2.2.
EXIST 2026 provides 3,984 training and 1,053 test memes balanced between English and Spanish. The Soft-Soft track treats each target as the empirical annotator distribution over labels, making the problem one of distribution prediction rather than point classification. Tasks 2.2 and 2.3 are semantically defined only when the sexism probability is non-zero, which motivates the hierarchical conditional treatment: downstream predictions are factored as P(Subtask)=p^(1)⋅p^(sub) at decoding time.
A distinctive element of EXIST 2026 is the provision of auxiliary physiological recordings—EEG bandpower features, gaze traces, and heart rate. The authors conduct a principled separability analysis restricted to EEG: after PCA to 21 components (~95% variance explained), MANOVA over majority-group soft labels shows no significant linear separability for Tasks 2.1 and 2.3, and although Task 2.2 reaches permutation significance (pperm=0.0010), the effect is negligible (Wilks' λ=0.9946, partial η2<0.01). No individual channel survives Bonferroni correction. This negative result justifies the decision to exclude physiological signals from the final model; notably, gaze and heart rate were excluded a priori rather than empirically tested, an assumption the authors acknowledge explicitly.
Architecture and training
The trainable pipeline is deliberately compact. A frozen Gemini Embedding 2 encoder produces a fixed 768-dimensional embedding per meme, which passes through a single SwiGLU block (expansion factor 2, dropout 0.2) followed by a residual projection. Three independent linear heads emit logits for the subtasks; the entire trainable portion comprises only 3.5M parameters. This design avoids fine-tuning large generative models on limited data while retaining rich multimodal semantics, following the efficiency rationale of fusion-based architectures such as Hate-CLIPper without its computational cost.
Training optimizes KL divergence against empirical soft targets for each head, with binary KL applied per category for Task 2.3. Task balancing uses homoscedastic uncertainty weighting with learnable variances σi, yielding self-paced learning in which high-loss tasks are down-weighted while the logσi regularizer prevents complete abandonment. Losses for downstream heads are masked to instances where downstream annotations exist, preventing non-sexist memes from inducing arbitrary gradient updates—an instance of hierarchical constraint enforcement through loss masking alone.
As an ablation (Run 3), the authors test detached soft-gating, scaling the shared representation by the stop-gradient of the predicted sexism probability before the downstream heads. The gating variant performs comparably to the base architecture locally and on the leaderboard, supporting the claim that conditional loss masking plus probabilistic decoding suffices to capture the hierarchy without architectural modification.
Inference proceeds via joint probability decoding for soft labels and naive thresholding at 0.5 for hard labels, with empty category sets resolved by maximum-probability selection to satisfy the ontological constraint that sexist memes carry at least one category. No threshold tuning was performed for the Hard-Hard track.
Results
Local evaluation on a 10% stratified split reveals substantial seed variance, particularly on downstream tasks—consistent with the small-data, high-disagreement setting. Run 1 led on Task 2.2 (ICM-S Nr 0.4367) while Run 2 led on Task 2.3 (ICM-S −3.2130); all runs were submitted to capture this variability.
Official Soft-Soft results:
| Task |
Best run |
Rank |
ICM-Soft |
ICM-Soft Norm |
| 2.1 |
aiwizards_3 |
4 |
0.2323 |
0.5373 |
| 2.2 |
aiwizards_1 |
4 |
−0.6720 |
0.4285 |
| 2.3 |
aiwizards_2 |
1 |
−2.8881 |
0.3469 |
Hard-Hard results are markedly weaker (ranks 12–19), obtained with untuned thresholds. The authors draw three observations: the system is better matched to soft-label prediction, for which it was designed; performance is strongest on the most difficult subtask (Task 2.3), where it placed first, second, and third across its three runs; and probabilistic outputs retain sufficient discriminative structure for non-trivial hard-label results even under naive thresholding. The Soft-Soft versus Hard-Hard gap indicates that distribution prediction and discrete decision quality are related but distinct objectives, with the latter likely benefiting from task-specific calibration—a concrete open question the paper leaves unresolved.
Limitations
The authors identify three principal limitations. First, dependence on proprietary Gemini Embedding 2 constrains full reproducibility, though the modular design permits substitution with open-weight backbones. Second, the fixed 0.5 threshold means Hard-Hard results should not be read as an upper bound. Third, the physiological analysis covered only EEG under linear methods; gaze and heart rate were excluded on dimensional and relevance grounds rather than measured evidence, so their utility—individually or in combination, or under non-linear modeling—remains untested. More broadly, the strong seed variance observed locally suggests that conclusions about architectural choices (e.g., the gating ablation) rest on comparisons within noise margins typical of this data regime.
Conclusion
The paper presents an efficient, disagreement-aware hierarchical multi-task framework for sexism identification in memes, demonstrating that frozen vision-language embeddings coupled with a 3.5M-parameter gated MLP, KL-based soft-label training, and learned uncertainty weighting can achieve top-ranked performance on the hardest subtask of EXIST 2026. The negative EEG separability result and the soft/hard evaluation gap delineate clear boundaries of the current approach, leaving open the questions of open-weight backbone substitution, calibrated hard-label decoding, and non-linear physiological fusion.