From Large to Small Language Models for Balanced Performance and Benchmarking

Large Language Models (LLMs) have transformed natural language processing, but their significant computational demands and environmental costs raise concerns. This study explores Small Language Models (SLMs) as a sustainable alternative, evaluating their efficiency and performance on the Massive Multitask Language Understanding (MMLU) benchmark. We assess three SLMs: TinyLlama 1.1B, Qwen2 1.5B, and Phi 3 Mini 3.8B, and introduce a novel hierarchical multi-agent system to enhance their capabilities. The findings demonstrate that SLMs achieve competitive baseline accuracy with notably lower computational overhead. Phi 3 Mini, for instance, reaches 69.0% accuracy while using only 7.6 GB of GPU memory, making it highly accessible. The hierarchical multi-agent system further bridges the performance gap with LLMs by providing specialized reasoning agents, tool-use capabilities, and a hierarchical consensus mechanism. This approach delivered substantial accuracy improvements: TinyLlama improved from 24.42% to 35.55%, Qwen2 from 51.44% to 61.39%, and Phi-3-Mini from 67.16% to 76.60%. The multi-agent paradigm offers significant environmental and efficiency benefits by improving models through intelligent composition rather than parameter scaling. This allows SLMs to achieve high performance with a drastically smaller footprint, reducing inference costs and promoting more sustainable AI. Our results support the use of SLMs as a viable and democratized alternative for AI development.