A TextGrad-Inspired Optimization Pipeline for AI-Assisted Generation of Lecture Notes and Problem Sets

Abstract

Large language models (LLMs) hold substantial promise for AI-assisted education, yet challenges such as hallucinations and pedagogical misalignment hinder their direct application. This article introduces a pipeline adapting textual gradient-based optimization for the automated generation and iterative refinement of lecture notes and problem sets. Addressing LLM limitations through a seven-step framework inspired by mathematical gradient descent, the proposed pipeline operates as an instructional copilot. Grounded in rubrics aligned with Bloom’s Revised Taxonomy, the framework utilizes course outlines and textbooks to ensure controllable content refinement, structural coherence, and cognitive engagement. Evaluated across a Science, Technology, Engineering, and Mathematics (STEM) course (“Numerical Methods”) and a non-STEM course (“Introduction to Macroeconomics”), the proposed pipeline significantly outperformed ablation variants (including zero-shot prompting and unguided iterative refinement) in educator-rated quality. Quantitative analyses demonstrate rapid self-correction, systematic generation of interactive elements, and robust component indispensability. A complementary student pilot study provided preliminary evidence of positive learning gains and user receptivity. This framework reduces educator workload and mitigates generative flaws, paving the way for scalable teaching tools while emphasizing the necessity of human oversight in AI-driven pedagogy.