Recommender systems heavily influence what users discover online, yet their decision logic often remains opaque. Counterfactual explanations (CFX) show which changes in a user's interaction history would alter a recommendation, offering a viable interpretability route. In practice, CFX typically requires many edits, producing long, low‑level lists that are hard for users to grasp. Existing evaluations focus on edit counts and overlook whether the explanations are coherent and actionable. We argue that this gap stems from a lack of user‑centered abstraction and limited qualitative assessment. To address it, we investigate large language models (LLM) as mediators that compress raw CFX edit lists into concise natural‑language narratives describing higher‑level preference patterns. We fine‑tune these mediators via Preference Optimization, targeting correctness, informativeness, and linguistic quality, and evaluate them with human‑calibrated metrics derived from separate labeled datasets. Experiments on the MovieLens‑1M dataset span multiple counterfactual generators and language models. Summaries achieve high overall linguistic quality, while the strongest semantic results appear with high‑success CFX methods such as ACCENT and SPINRec. Direct Preference Optimization improves semantic metrics in almost all settings.\n\nReview: The study highlights the promise of LLMs in translating technical counterfactual edits into user‑friendly narratives, advancing explainability for recommender systems.