Making AI Think and Reason Like a Senior Developer: Infusing Tacit Knowledge into LLMs

Abstract

Large Language Models (LLM) are transforming software engineering in a rapid manner, supporting automatic code generation, debugging, and architectural proposals.Even now, with these developments, AI systems lack the sophistication of the reasoning of senior developers, whose judgments are highly affected by tacit knowledge derived by experience. Tacit knowledge encompasses contextual judgment, trade-off reasoning and heuristic-based problem solving that are difficult to capture in the conventional datasets that are utilized to train AI models. This review has investigated the new avenue of research of instilling tacit knowledge in LLMs so that they can reason at the expert level when developing AI-assisted software development. The paper has addressed the conceptual underpinnings of tacit knowledge, shortcomings of current LLM knowledge capabilities, knowledge-enhanced architectures, and a proposed theoretical model of Tacit Knowledge-augmented LLM. The conceptual models and experimental interpretations raised the point that incorporating the context of the workflow and loops of expert feedback can enhance reliability and the depth of reasoning. The paper concludes that reconciling experiential knowledge with data-driven learning is a milestone toward the development of next-generation AI systems that will be able to act as collaborative engineering partners.