Tag
The paper presents a nine-voter ensemble system using error-independent LLMs for harmful content detection in German social media, achieving first place in GermEval 2026 shared task across four subtasks by addressing class imbalance.