In the context of the JOKER 2024 Task 2 Challenge, this paper presents an emerging approach that leverages the latent representations derived from different Large Language Models (LLMs) to drive a classification mechanism. Our methodology involves exploiting the "knowledge" encoded in LLMs to effectively discriminate humor genres. Experimental results show promising results, demonstrating the effectiveness of our approach. However, inherent complexities remain, such as the proximity between certain classes and biases arising from the dataset distributions. These complexities warrant further investigation to refine the classification process and improve overall performance.