We describe the Eraserhead system submitted to the PsyDefDetect shared task at BioNLP 2026, which frames psychological defense level detection as a nine-class utterance classification problem over supportive dialogue. Our system is based on Qwen3-14B and combines clinically informed prompt design, per-label oversampling, and careful inference settings for stable prediction. A central challenge of the task is strong class imbalance, with High-Adaptive responses appearing far more often than several minority classes. This makes it easy for models to favor the majority class and achieve reasonable accuracy while performing poorly on rarer categories. To address this, we iteratively adjusted oversampling targets based on error analysis and predicted label distributions across submission rounds. Our final system achieved an official macro F1 of 0.3418 on Leaderboard 1 and 0.3947 on Leaderboard 2, ranking 7th among the 21 registered teams on both leaderboards. We further analyze the main failure modes of the system, especially the difficulty of distinguishing Minor Image Distorting defenses from High-Adaptive responses and the persistent tendency to over-predict the majority class. These findings highlight the broader difficulty of modeling psychological function from text alone.

Memes and other text-embedded images are powerful tools for expressing opinions and identities, especially within marginalized socio-political movements. Detecting hate speech in this type of multimodal content is challenging because of the subtle ways text and visuals interact. In this paper, we describe our approach for Subtask A of the Shared Task on Multimodal Hate Detection in Marginalized Movement@CASE 2025, which focuses on classifying memes as either Hate or No Hate. We tested both unimodal and multimodal setups, using models like DistilBERT, HateBERT, Vision Transformer, and Swin Transformer. Our best system is the large multimodal model Qwen2.5-VL-7B-Instruct-bnb-4bit, fine-tuned with 4-bit quantization and instruction prompts. While we also tried late fusion with multiple transformers, Qwen performed better at capturing text-image interactions in memes. This LLM-based approach reached the highest F1-score of 0.8086 on the test set, ranking our team 5th overall in the task. These results show the value of late fusion and instruction-tuned LLMs for tackling complex hate speech in socio-political memes.