CLAIRE: A Controllable LLM Tutoring Framework for Reading Comprehension

Reading comprehension requires strategic and metacognitive support provided by guided reading methodologies through structured, scaffolded interaction. However, enacting these methodologies systematically at scale remains challenging, particularly in digital contexts. Recent advances in Large Language Models (LLMs) open new opportunities, though existing approaches largely rely on unstructured dialogue and implicit pedagogical control. In this paper, we present CLAIRE, a teacher-configurable framework that operationalizes guided reading by decomposing the pedagogical procedure into four components, i.e., phases, interaction moves, assessment criteria, and transition logic, authored by teachers and executed at runtime through a structured multi-agent dialogue controller. Experimental evaluation shows that CLAIRE can faithfully enact a teacher-defined guided reading methodology and is perceived by teachers as pedagogically meaningful and useful, supporting the scalable adoption of structured reading instruction with LLMs. Code and data are available at https://github.com/softlab-unimore/Claire.