Small Language Model Enhancement Strategies in Practice: A Signal-Oriented Taxonomy and Open Questions (Extended Abstract)

Abstract

Deploying small language models (SLMs) for strategic decision-making in regulated domains requires balancing accuracy, cost, and auditability. Fine-tuning, cascading, and prompt augmentation can improve performance, but these approaches are typically studied independently. This paper presents a taxonomy that organizes enhancement strategies by the type of signal delivered to the SLM. Comparing strategies along this dimension shows differences not only in training requirements but also in what information crosses the model boundary at inference time. These distinctions have practical consequences for cost, auditability, and robustness that remain underexamined. The paper concludes with open questions for enterprise deployment.