Large Language Models (LLMs) have made waves in various fields, but can they improve recommendations? Traditionally, collaborative filtering methods, which analyze user-item interactions, have been the backbone of recommender systems. However, LLMs offer a new avenue by leveraging their extensive knowledge and language processing capabilities. The challenge lies in effectively integrating collaborative filtering information into LLMs. Existing methods often inject collaborative features directly into the input, which can disrupt the original text and limit the LLM’s ability to understand and infer correctly. A new approach called CoRA, or Collaborative LoRA, proposes a clever solution: aligning collaborative information with the LLM’s parameter space, essentially converting the collaborative information into additional weights that guide the LLM’s recommendations. This technique enables the LLM to absorb collaborative signals without hindering its general knowledge and text comprehension abilities. CoRA works by employing a collaborative filtering model to capture user-item relationships as embeddings. These embeddings are then converted into tailored weights that enhance the LLM’s understanding of individual user preferences. Experiments show that CoRA outperforms traditional methods, especially in scenarios where collaborative filtering information is crucial, such as recommending items to existing users. The innovation lies in its parameter space alignment, preventing interference with the LLM's textual understanding. This allows CoRA to excel in combining the LLM’s linguistic prowess with personalized insights from collaborative filtering. By working *with* the LLM's existing strengths, CoRA offers a promising new path for LLM-based recommendations.