Recommender systems play a crucial role in information filtering in the age of Big Data. In recent years, the paradigm of conversational recommender systems has emerged to enhance user interactivity and improve the overall recommendation experience. A fundamental capability of these systems is generating coherent and relevant responses to users. Recently, generative large language models (LLMs) have demonstrated a remarkable capacity for producing human-like responses, which can be effective in this context. However, these models face two significant challenges: the hallucination problem, which results in the generation of inaccurate or misleading information, and their substantial size. In this study, we investigate the performance of two instruction-tuned models, Mistral-7B and Llama3-8B, while also examining the impact of Parameter-Efficient Fine-Tuning (PEFT). Our experiments on a publicly available dataset show that PEFT, combined with a custom prompting template, can effectively guide LLMs to generate improved responses that align better with the intended actions of conversational recommenders.