Small Reasoning Language Models in Agentic Systems (SMART)

Recent research challenges the “bigger is better” paradigm in language models, demonstrating that appropriately optimized small language models (SLMs) can match or exceed larger counterparts on specific tasks. The successful distillation of reasoning abilities to smaller architectures suggests effective reasoning requires substantially less parameters than previously assumed. While extensive research has been conducted into SLMs and methods for enhancing reasoning capabilities in smaller language models, a research gap remains concerning the minimal parameter thresholds required to maintain sufficient reasoning abilities. This project plans to address this gap, particularly within agentic AI systems where specialized, repetitive tasks favor compact models.