A Neuro-Symbolic AI Approach for a Conversational Workout Plan Recommender System

Abstract

Not being active enough is a global health problem, and the fitness apps that are out there at the moment don’t give guidance that’s personalized, safe and interesting enough. The thing with recommender systems at the moment is that they’re either really secure but a bit boring, or fun but a bit unreliable. This paper proposes and evaluates a neuro-symbolic framework embedding the logical consistency of ontology-based reasoning into the conversational fluency of large language models. At the core of this research is a controlled comparative experiment between a baseline model, Gemma3-Ground, and the proposed model, Gemma3-Chase, finetuned on a corpus containing explicit, inferred reasoning traces generated by SWRL rules. A multi-faceted evaluation performed using 20 synthetic user profiles showed that Gemma3- Chase consistently outperformed Gemma3-Ground significantly across key metrics, especially on effectiveness, semantic relevance, and groundedness. The neuro-symbolic model also had a much smaller standard deviation, pointing to its high reliability. The study validates the embedding of logical reasoning into the fine-tuning process of LLM as an effective method of constructing conversational recommender systems that are trustworthy and engaging for the health and wellness domain.