Abul Hasnat

RL-based post-training with Group Relative Policy Optimization (GRPO) can significantly boost the ability of thinking-based MLLMs to detect hateful memes by improving both classification accuracy and explanation quality.