Pedagogical Alignment of Large Language Models

Large Language Models (LLMs) are an integral part of the modern digital landscape, and ensuring their alignment to be honest, harmless, and helpful - the '3H' principle - is critical to their effective usage [1]. However, the needs of a student learning a subject or preparing for an exam differ substantially from those of a casual internet user. While the latter might seek immediate answers to their queries, the former benefits more from a structured, guided approach that encourages active learning [2] and critical thinking [3]. Therefore, if we aim to leverage LLMs to assist students' learning, we must redefine the'helpfulness' criterion in a way that specifically supports this demographic [4]. In this paper, we propose a novel concept of pedagogical alignment for LLMs which aligns LLMs to function as scaffolding tools [5], breaking complex problems into manageable subproblems and guiding students towards the final answer through constructive feedback and hint [6]. This approach redefines the'helpfulness' criterion of the '3H' principle, tailoring it to the unique needs and learning objectives of students. We argue that to truly assist students, LLMs should not merely provide immediate answers, but rather guide learners through a structured, step-by-step process that encourages active learning and critical thinking.