Pedagogical Alignment of LLMs requires Diverse Cognitively-Inspired Student Proxies

Large Language Models (LLMs) are increasingly positioned as tutors, judges, and instructional assistants. Yet their pedagogical performance remains shallow: they optimize for producing correct answers rather than for teaching. Pedagogy requires anticipating misconceptions, sequencing curricula, calibrating task difficulty, and adapting interactively to learner trajectories. In this position paper, we characterise the ideal behaviour of LLMs assuming a Teacher Role (pedagogical agents). Teaching involves selecting informative examples and learning from them as a basis for inference about what demonstrations a helpful teacher would provide. Given this, we argue that the pedagogical capabilities of LLMs as Teachers are limited due to the limited meta-reasoning capabilities of LLMs. We motivate structured cognitively-inspired student proxies as indispensable for pedagogical alignment. Student proxies are constrained cognitive models that generate structured, interpretable error trajectories. We argue that a Teacher–Student Agentic Framework (TSAF) with heterogeneous student proxies can enable teacher LLMs to improve their pedagogical alignment – by adapting reasoning and feedback strategies, monitoring errors, and scaffolding learning efficiently across tasks. Our position is that cognitive proxies reframe pedagogical alignment as an interpretable, principled paradigm for enhancing how LLMs “learn to teach”