From Bias to Fairness: Mitigating Popularity Bias in LLM-Based Recommendation Systems via Supervised Fine-Tuning and Direct Preference Optimization

Large Language Models (LLMs) have demonstrated strong potential as recommendation engines, leveraging their broad world knowledge and natural language understanding to generate personalized item suggestions. However, LLMs inherit and amplify popularity bias present in their pretraining corpora, consistently recommending well-known items while neglecting the vast majority of the catalog. This paper presents a systematic study of popularity bias across four inference and training regimes applied to Qwen2.5 models of two scales (1.5B and 3B parameters) on the MovieLens-100K benchmark. We show that zero-shot and few-shot prompting produce heavily skewed recommendations, and that naive Supervised Fine-Tuning (SFT) inherits the dataset's popularity distribution (71.4\% head items in raw training pairs). We introduce a popularity-capped sampling strategy that rebalances SFT training data to 73.9\% tail items, substantially improving fairness. Building on this, we apply Direct Preference Optimization (DPO) using preference pairs constructed from user ratings, where long-tail liked items serve as chosen responses and popular disliked items serve as rejected responses. Our best configuration (Qwen2.5-3B, SFT + DPO) achieves a 59.2\% reduction in Average Recommendation Popularity (ARP) and a Tail Coverage (TailCov) of 95.45\%, compared to 37.21\% for the zero-shot baseline. These results demonstrate that alignment-stage training is a powerful and practical tool for fairness in LLM-based recommendation.