ELiRF-UPV at EXIST 2026: Multimodal Fusion and Few-Shot LLMs for Sexism Detection in Memes and Videos

This paper describes the participation of ELiRF-UPV in Tasks 2 and 3 of EXIST 2026 at CLEF 2026, covering all six subtasks of sexism identification, source-intention detection, and fine-grained sexism categorization in memes and TikTok videos. Following the official protocol, we report ICM and F1 in the hard–hard setting and ICM-Soft in the primary soft–soft setting, as computed by PyEvALL. We submit two complementary families of systems, both applied to memes and videos and both relying on neutral Qwen3-VL-8B captioning to derive a visual description and an independent gender-relevance analysis that avoids confirmatory bias. The first, PhysioMeme-Fusion, is a trained multimodal system with three per-subtask heads optimized directly on soft labels—KL divergence for the mono-label heads and binary cross-entropy for the multi-label categorization—each with an explicit “No” class so that the soft metric drives learning. Its core pairs language-specific emotion-tuned encoders for English and Spanish with an enriched textual view of each instance (the on-screen text combined with the neutral visual description and the gender-relevance analysis, denoted capsanocr), and adds a physiological branch encoding the eye-tracking, heart-rate, and EEG signals released for the lab subjects, fused through a residual cross-attention block with modality dropout; an optional dense visual stream—SigLIP2 for memes and a VideoPrism video encoder for videos—is explored as an add-on. The second family is a training-free pipeline that prompts Qwen3.5-27B and Gemma-4-31B in a hierarchical cascade, passing meme images or sampled video frames directly to the model alongside the released text. Our strongest result is the training-free few-shot pipeline: under the hard–hard evaluation it is competitive across both modalities and, in the official by-team ranking, places first among teams on video identification (T3.1), second on meme categorization (T2.3), and third on video intention (T3.2), which shows that a definition-rich hierarchical prompt is a powerful, data-free baseline for multimodal sexism detection. On the trained side, the soft-optimized fusion reaches the upper part of the primary soft–soft leaderboard on meme identification (T2.1, seventh among teams), and we find that the enriched textual view—building on the VLM-derived enrichment of Arcos et al.—carries the system, whereas the physiological branch yields no soft-metric gain (and is at times detrimental on the fine-grained categorization) and the dense visual stream degrades performance wherever the verbalized view is already present.