一句话：研究提示形式与认知启发推理框架，提升小语言模型推理能力，推动AI能力民主化。
AI 正在重塑软件工程，但能力最强的工具往往难以被广泛研究者与组织可靠使用：一些仅提供封闭接口且变化不可预测，损害可复现性与长期科学进展；另一些要求普通学术实验室或工业团队难以承担的硬件，限制了参与门槛。小语言模型提供了实用折中：开放、高效，并可在受控环境中部署，但在需要深度推理的任务上仍落后于大模型。本论文旨在弥合这一差距，研究提示的形式与结构如何影响模型表现，并受人类认知原理启发开发新的推理框架。通过增强小模型的问题求解能力，本工作支持一个先进 AI 能力可访问、可审计且可被更广泛软件工程社区使用的未来。
AI is reshaping Software Engineering, but the tools with the highest capabilities remain difficult for many researchers and organizations to rely on. Some are accessible only through closed interfaces that change unpredictably, hindering reproducibility and long-term scientific progress. Others require hardware far beyond the reach of typical academic labs or industry teams, limiting who can meaningfully participate in this technological shift. Small Language Models offer a practical middle ground: they are open, efficient, and deployable in controlled environments. Yet they continue to lag behind larger models in tasks requiring deep reasoning. This dissertation aims to bridge that gap by studying how the form and structure of prompts influence model performance, and by developing new reasoning frameworks inspired by principles of human cognition. By enhancing the problem-solving abilities of smaller models, this work supports a future in which advanced AI capabilities are accessible, auditable, and usable by the broader SE community.