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

AI summary

Overview

Research area: Artificial Intelligence — LLM-based multi-agent systems, procedural video generation, and AI for education (K–12 instructional media).

Technical level: Advanced. The paper assumes familiarity with large language models, multi-agent orchestration, code generation, and deterministic rendering pipelines (Manim).

Scope: LASEV is a hierarchical LLM multi-agent system that generates instructional videos by producing an executable, editable script (rather than pixels), which is then deterministically compiled into synchronized visuals and narration.

What This Paper Is About

End-to-end text-to-video models such as Sora and the Wan series produce visually convincing output but operate in probabilistic pixel space, which makes them unreliable for content that demands strict logical and symbolic accuracy, such as K–12 problem-solving videos. The paper argues that educational video generation is not a pixel-prediction problem but an orchestration problem over reasoning, symbolic content, and timing. To solve it, the authors propose LASEV, a system in which specialized LLM agents collaborate to build a structured executable video script that is compiled deterministically into a finished instructional video.

Key Contributions

1. A structured LLM-based multi-agent system for instructional video generation. Complex pedagogical workflows are decomposed into coordinated reasoning, visualization, and narration stages under a single central Orchestrating Agent, which owns process planning and the final alignment specification.

2. A heterogeneous critique and verification mechanism. Every intermediate artifact is gated by three independent checks — semantic rubrics for pedagogical correctness, tool-based execution (Python/Manim compilation) for functional feasibility, and rule-based constraints for structural compliance — feeding an iterative critique–revision loop until quality gates are satisfied.

3. The Executable Video Script (EVS) formulation. Video is reframed as a triplet S = (P, N, A), where P is pedagogical content (premises, step-by-step reasoning, symbolic assets such as LaTeX equations and geometric constructs), N is a temporal sequence of narration segments, and A is the programmatic alignment and orchestration rule set. The EVS is compiled through V = Compile(S) = Render_vis(P, A) ∥ Synth_audio(N).

4. An end-to-end automated production pipeline. Multi-agent collaboration, structured scripting, and executable visualization code are unified into a template-driven assembly process that requires no manual editing and is validated at industrial scale.

Main Findings

• Instructional quality gains over direct prompting. On Elementary Chinese Language Arts, LASEV reaches 96.0% Usable Rate, 92.0% Publishable Rate, and 4.0% Perfect Rate, versus Qwen3 at 84.0/82.0/6.0, DeepSeek-R1 at 80.0/70.0/5.0, and GPT-4o at 74.0/52.0/0.0.

• State-of-the-art results on Middle School Mathematics. LASEV achieves 96.0% Usable Rate, 96.0% Publishable Rate, and 58.0% Perfect Rate. Baselines range from DeepSeek-R1 with basic prompting (45.5/45.5/27.3) to GPT-4o with improved prompting (94.0/86.0/52.0).

• Multi-agent collaboration is the largest single contributor. The single-agent baseline reaches only 42.0% Publishable Rate on Chinese and 60.0% Usable Rate on Math. Adding Semantic Critique alone lifts Chinese Publishable Rate by 40.0 points (52.0% → 92.0%), the largest gain of any component.

• Each critique type contributes measurably. Tool-based Critique adds 22.2 points to Chinese Publishable Rate by catching Python runtime and Manim compilation failures; Rule-based Critique adds 10.0 points through format and keyword compliance checking. Removing any single component reduces the Math Perfect Rate by at least 10 percentage points.

• A documented quality trade-off on Chinese. The Full System's Perfect Rate (4.0%) is lower than some ablated variants (e.g., 10.0% without Rule-based Critique) because strict structural compliance pushes borderline cases from Score 1 to Score 2 (82.0% → 92.0% Publishable Rate), occasionally introducing minor rephrasing artifacts scored as 2 rather than 3.

• Cost and throughput at industrial scale. LASEV's full system costs ≤ $0.07 (0.5 CNY) for a 2-minute video with a 96% usable rate and capacity of 1M videos/day, versus $1.40 (10 CNY) for the Education Industry Standard at 95% usable rate and 30K/day capacity — described as over 95% cost reduction. The LLM component alone costs $0.056 (0.4 CNY) and the optional digital human ≤ $0.014 (0.1 CNY).

• Errors are mostly cosmetic, not pedagogical. In Math, minor issues make up 38.0% of outputs (presentation flaws 22.0%, minor redundancy 6.0%, imprecise description 4.0%, minor step skipping 2.0%, non-optimal method 2.0%, auxiliary line issues 2.0%). Critical issues total only 4.0%, split evenly between solution errors (2.0%) and problem misreading (2.0%). Total non-perfect rate is 42.0%.

• Production deployment reliability. Production logs covering over 20 million generated videos show a 3.58% overall failure-or-filtering rate; 96.42% are successfully produced. Policy-based filtering accounts for 1.30%, solution/script generation failures for 1.32%, and animation/LaTeX compilation failures for 0.90%.

• Evaluation robustness. Each video is scored independently by three certified domain experts with at least one year of teaching experience; all three experts assign identical scores to 85% of evaluated videos. The offline evaluation set contains 493 Chinese Language Arts problems and 510 Mathematics problems, each paired with one generated instructional video.

Methodology in Plain English

The authors begin by diagnosing why pixel-based video models fail at education: they approximate how equations and diagrams look rather than execute their meaning, so they produce distorted equations, inconsistent diagrams, and flawed logical progressions. The fix is to change the output space from pixels to a script.

The workflow runs as follows. A central Orchestrating Agent receives an educational problem, decomposes it into subtasks, and decides whether a dynamic illustration is warranted. It then dispatches work to three specialized agents. The Solution Agent produces a step-by-step reasoning trace organized as logically atomic steps. The Illustration Agent writes actual Python/Manim code that draws the visuals, rather than describing them. The Narration Agent writes learner-friendly explanatory text segmented so each segment maps to a discrete logical unit in the solution.

Nothing is accepted on first pass. Each artifact passes through three gates: an LLM-based semantic review against rubrics and curriculum-aligned syllabi (with few-shot examples embedded in the critique prompt to stabilize judgment), a deterministic tool check that actually compiles and runs the visualization code to catch hard errors, and a rule-based audit for keyword matching, format, and API constraints. Failures return structured feedback to the responsible agent, and the loop repeats for up to 3 refinement iterations per stage with DeepSeek-R1 as the backbone model.

Once content and narration are validated, the Orchestrating Agent alone writes the alignment layer — temporal mapping between narration segments and visual events, style constraints such as fonts, layout and transitions (mostly predefined in templates), and scene evolution logic. This is deliberately not delegated to any working agent, so end-to-end consistency is guaranteed at compilation. Templates expose controllable slots for content and narration; display duration is driven by symbolic density; animation clips are inserted with synchronized triggers when illustration is present. The script is then compiled deterministically into a video with synchronized audio.

Why This Matters

Impact on research. The paper reframes a generation problem as a compilation problem, arguing that the right abstraction for domains with strict symbolic requirements is an executable intermediate representation rather than a latent pixel distribution. It also contributes a heterogeneous verification design — combining stochastic LLM critique with deterministic compilation and rule checks — that is directly relevant to any multi-agent pipeline where outputs must be functionally correct rather than merely plausible.

Real-world applications:

• Automated production of K–12 mathematics and language arts instructional videos, including step-by-step worked solutions for algebra, geometry, and word problems.
• Large-scale educational content platforms that need millions of videos per day at per-video costs below $0.07.
• Tutoring and homework-help products where an explanation must be mathematically exact and cannot tolerate a hallucinated diagram or distorted equation.
• Curriculum-aligned content pipelines where rule-based constraints enforce institutional formatting, notation, and API standards automatically.

Industry relevance. The cost table positions LASEV against both traditional education production and general video generators (Runway Gen-3, Sora 2 Standard and Pro, Kling AI, Hailuo AI, Vidu). The reported combination of a 96% usable rate, 1M videos/day capacity, and roughly 95% cost reduction relative to the education industry standard targets a gap that neither high-quality/high-cost nor low-cost/low-accuracy approaches fill. The reported 3.58% failure rate across over 20 million production videos indicates the system has been run as a real service, not only as a benchmark.

Future Directions

• Improving the Perfect Rate on language arts content. LASEV's 4.0% Perfect Rate on Chinese is below several baselines and below some of its own ablations, so the rephrasing artifacts introduced by strict rule-based revisions remain an unresolved quality issue.

• Reducing residual failure modes. Production failures cluster at 1.32% for solution/script generation and 0.90% for animation/LaTeX compilation; the paper attributes these to bounded local generation and tool-execution issues, leaving room for more robust code generation and repair.

• Extending beyond the evaluated domains. Only Elementary Chinese Language Arts and Middle School Mathematics are evaluated, so generalization to other subjects, grade levels, or higher education is not established.

• Integrating the digital human module. The paper notes that LASEV may optionally integrate a digital human via an independent module, which it explicitly places outside the scope of this work.

• Scaling the illustration decision. The Illustration Agent is optional and the orchestrator decides case-by-case whether dynamic visualization reduces cognitive load; how that decision generalizes across problem types is left open.

Target Audience

This paper is most valuable to researchers and engineers building multi-agent LLM systems or agentic content pipelines, to AI-for-education teams and edtech companies weighing the cost and reliability trade-offs of automated video production, and to practitioners working on code-generation-as-visualization approaches that use Manim or similar rendering frameworks. It will also interest anyone studying verification and critique mechanisms in agentic workflows, since the heterogeneous critique design and the production failure statistics are the paper's most transferable engineering contributions.

Authors’ abstract

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 LASEV, a hierarchical LLM-based multi-agent system for generating high-quality instructional videos from educational problems. LASEV 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--LASEV decomposes the generation workflow into specialized agents that collaborate through a central Orchestrating Agent, shared production state, 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 code, 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 production without manual editing. In large-scale deployments, LASEV 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.