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

Although recent end-to-end video generation models demonstrate impressive performance in visually oriented content creation, they remain limited in scenarios that require strict logical rigor and precise knowledge representation, such as instructional and educational media. To address this problem, we propose LAVES, a hierarchical LLM-based multi-agent system for generating high-quality instructional videos from educational problems. The LAVES formulates educational video generation as a multi-objective task that simultaneously demands correct step-by-step reasoning, pedagogically coherent narration, semantically faithful visual demonstrations, and precise audio–visual alignment. To address the limitations of prior approaches–including low procedural fidelity, high production cost, and limited controllability–LAVES decomposes the generation workflow into specialized agents coordinated by a central Orchestrating Agent with explicit quality gates and iterative critique mechanisms. Specifically, the Orchestrating Agent supervises a Solution Agent for rigorous problem solving, an Illustration Agent that produces executable visualization codes, and a Narration Agent for learner-oriented instructional scripts. In addition, all outputs from the working agents are subject to semantic critique, rule-based constraints, and tool-based compilation checks. Rather than directly synthesizing pixels, the system constructs a structured executable video script that is deterministically compiled into synchronized visuals and narration using template-driven assembly rules, enabling fully automated end-to-end production without manual editing. In large-scale deployments, LAVES achieves a throughput exceeding one million videos per day, delivering over a 95% reduction in cost compared to current industry-standard approaches while maintaining a high acceptance rate.

💡 Research Summary

The paper introduces LAVES (Large‑Language‑Model‑Based Multi‑Agent System for Educational Video), a hierarchical, LLM‑driven framework that automatically generates high‑quality instructional videos from educational problem statements. Recognizing the fundamental mismatch between conventional text‑to‑video diffusion models—which predict pixel distributions—and the strict logical, symbolic, and temporal precision required for K‑12 mathematics and science instruction, the authors reformulate video generation as the creation of an Executable Video Script (EVS). An EVS is a triplet S = (P, N, A) where P contains the pedagogical content (problem description, step‑by‑step solution, and symbolic visual definitions), N is a temporally ordered narration aligned with each logical step, and A encodes alignment rules that synchronize visuals and audio at the frame level.

The system architecture consists of a central Orchestrating Agent and three specialized Working Agents: a Solution Agent that uses a large language model to produce rigorous, step‑wise reasoning and the textual component P; an Illustration Agent that generates executable visualization code (e.g., Python/Manim scripts) rather than raw pixels, guaranteeing mathematically exact animations; and a Narration Agent that writes learner‑oriented explanatory text N and passes it to a Text‑to‑Speech module. The Orchestrating Agent parses the input problem Q, maintains global production state, and defines the alignment specification A.

Every output passes through a three‑layer critique pipeline. First, a semantic reviewer (LLM‑based) checks pedagogical correctness, terminology consistency, and logical flow. Second, rule‑based validators enforce structural constraints such as allowed functions, keyword usage, and formatting. Third, tool‑based checks compile and render the visualization code and synthesize the audio, ensuring functional feasibility. If any check fails, detailed feedback is fed back to the responsible agent, triggering an iterative critique‑revision loop until all constraints are satisfied.

Once a validated EVS is obtained, a deterministic rendering pipeline compiles it into the final video V: Render_vis interprets P and A to produce a precise visual stream, while Synth_audio converts N into speech. The two streams are merged under the strict temporal mapping defined in A, achieving frame‑accurate synchronization between narration and visual events.

Large‑scale experiments simulate production of over one million videos per day, reporting a cost reduction exceeding 95 % compared with current industry pipelines and a high acceptance rate from downstream educators. The code‑based visual generation eliminates common diffusion‑model failures such as distorted equations, inconsistent diagrams, and logical discontinuities.

In summary, LAVES demonstrates that educational video synthesis can be reframed from stochastic pixel generation to structured script synthesis, leveraging LLMs for reasoning, code generation, and language production while enforcing multi‑modal consistency through heterogeneous critiques. This yields a controllable, scalable, and cost‑effective solution that meets the rigorous demands of instructional content. Future work is suggested on extending the framework to broader curricula, incorporating personalized feedback loops, and enabling real‑time interactive educational experiences.