A new approach to emotion recognition in speechLLMs improves Macro F1 and reduces hallucinations by adapting the model to read the final prompt token's hidden state through a classification head, providing controlled comparison of generative and discriminative inference.