Mitigating Hallucinations in Small Language Models: A Performance Analysis of DoLa Optimization for Mobile Devices

As the demand for efficient and reliable natural language processing solutions on mobile devices grows, Small Language Models (SLMs) have gained prominence. This paper analyzes various small language models, focusing on their performance concerning hallucination mitigation techniques, specifically Decoding by Contrasting Layers (DoLa). SLMs, while efficient, often face challenges in balancing performance and resource constraints. By systematically evaluating the effectiveness of DoLa across different such architectures, we provide insights into how this technique can enhance the reliability of small language models while maintaining their efficiency for mobile applications.