Title: Beyond End-to-End Video Models: An LLM-Based Multi-Agent System for Educational Video Generation

Abstract:

While contemporary end-to-end video generation models have shown remarkable capabilities in creating visually driven content, they often fall short in contexts demanding strict logical precision and accurate knowledge representation, such as instructional and educational media. To bridge this gap, we introduce LASEV, a hierarchical, multi-agent system powered by Large Language Models (LLMs) designed to produce high-quality instructional videos from educational problems. LASEV treats video generation as a complex, multi-objective challenge that requires simultaneous adherence to correct step-by-step reasoning, pedagogically sound narration, semantically accurate visual demonstrations, and precise audio-visual synchronization.

To overcome the drawbacks of previous methods—namely low procedural fidelity, excessive production costs, and limited controllability—LASEV breaks down the workflow into specialized agents that collaborate via a central Orchestrating Agent. This system utilizes a shared production state, explicit quality gates, and iterative critique mechanisms to ensure reliability. Specifically, the Orchestrating Agent oversees three key components: a Solution Agent dedicated to rigorous problem-solving, an Illustration Agent responsible for generating executable visualization code, and a Narration Agent tasked with creating learner-focused instructional scripts.

Furthermore, all outputs from these working agents undergo rigorous evaluation, including semantic critique, rule-based constraint checking, and tool-based compilation verification. Instead of directly synthesizing pixels, LASEV constructs a structured, executable video script that is deterministically compiled into synchronized visuals and narration through template-driven assembly rules. This approach facilitates fully automated production without the need for manual editing. In large-scale deployments, LASEV demonstrates exceptional efficiency, achieving a throughput of more than one million videos daily. This performance represents a cost reduction of over 95% compared to current industry-standard methods, all while maintaining a high acceptance rate.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC