An AI-First Proof of Concept: Simulating and Refining a Teach-Back Protocol for Dialogic Learning in Programming Education

Marta Valentini

Keywords: Generative AI agents, AI-first research workflow, Teach-Back protocol, Dialogic tutoring, Introductory programming education (CS1), Metacognitive scaffolding

TL;DR: AI-first dialogic Teach-Back tutoring in CS1: a proof-of-concept with simulated sessions that expose misconceptions and refine the protocol.

Abstract:

This AI-first proof-of-concept investigates whether a generative agent can design
and enact a dialogic Teach-Back protocol for novice programmers. We ran three
controlled simulations on a C++ for-loop task and conducted a lightweight thematic
analysis of the resulting dialogues. The analysis surfaced recurrent misconceptions
and metacognitive moves and informed three refinements to the protocol: a
dedicated Slow-Thinking phase, incremental probing to manage cognitive load,
and a Transfer phase to prompt forward-looking application and self-explanation.
Findings are illustrative from AI self-play and bounded to this simulated setup;
they indicate how an AI can act as a dialogic partner—rather than an answer engine—
supporting conceptual clarification and metacognitive reflection. We release
prompts and anonymized logs to enable inspection and reuse. A classroom deployment
in Programming 1 is planned, positioning the protocol as a transparent,
method-driven template for integrating AI into learning environments.

Submission Number: 212