All projects
research2026Multimodal AILLMs

GEMF — EXIST 2026

Multimodal AI for sexism characterization in memes

GEMF — EXIST 2026ImageOCRGeminiEncoderFusionp(y)
GEMF — EXIST 2026Multimodal system
3× #1Official first places
1 / 144Task 2.1 soft rank
0.741 → 0.880Validation AUC

01

Problem

Memes are not separable image and text signals: meaning often lives in the relationship between both, while labels also reflect genuine annotator disagreement.

02

Approach

Gemini translates visual and pragmatic meaning into structured text instead of deciding the final class. Supervised encoders fuse that representation with OCR, EEG and Ekman emotion features.

03

Result

The system achieved three official first places across EXIST 2026 Task 2 leaderboards, while the repository also documents the hard-label failure mode in Task 2.3.

Transparency

Limitations

Task 2.3 hard underperformed because category thresholds did not transfer reliably to the test set.

Evidence

Sources

Next case

UrbanFlow Valencia