Accepted to Findings of ACL 2026

mPresenter: An Agentic Framework for Generating Multilingual Presentation Videos from Scientific Papers

An end-to-end multilingual Paper2Video framework that turns academic PDFs into narrated presentation videos through multi-agent planning, layout-aware slide generation, and cross-lingual figure interpretation.

Wenhan Han1, Xiao Xiao2, Mykola Pechenizkiy1, Meng Fang2,1
1 Eindhoven University of Technology
2 University of Liverpool
w.han@tue.nl · xiao.xiao@liverpool.ac.uk · m.pechenizkiy@tue.nl · Meng.Fang@liverpool.ac.uk
Code Benchmark Demo Videos

Abstract

Multilingual scientific presentations, generated end to end

mPresenter decomposes paper-to-video generation into specialized agentic stages for planning, critique, slide coding, visual inspection, figure interpretation, and video synthesis.

Generating presentation videos from scientific papers is challenging because systems must reason over long-document discourse, translate dense findings into slide-level narratives, preserve layout readability, and explain visual evidence across languages. Existing Paper2Video systems are largely monolingual and often rely on single-pass generation. mPresenter addresses these limitations with a multi-agent framework that produces Beamer LaTeX slides, visually inspects rendered outputs, augments narration with cross-lingual figure interpretation, and synthesizes final presentation videos. The project also introduces mPreBench, an expert-curated multilingual benchmark for evaluating Effective Information Transfer through question answering.

Framework

From PDF input to presentation video

The pipeline extracts paper assets, plans an outline and talk script, refines slides through compiler and readability feedback, interprets figures across languages, and synthesizes narrated video.

Overview of the mPresenter agentic pipeline
mPresenter orchestrates Planner, Reviewer, Coder, and Interpreter agents to produce coherent multilingual presentation videos.

Multi-agent collaboration

Planner, Reviewer, Coder, and Interpreter agents coordinate outline design, critique, slide creation, and visual explanation.

Cross-lingual interpretation

The Interpreter Agent analyzes figures and explains visual elements that remain in the source language.

True-to-layout generation

Executable Beamer LaTeX is compiled and visually inspected to improve readability and slide fidelity.

High information density

The system is designed for effective knowledge transfer from scientific papers to presentation videos.

Cost-effective efficiency

mPresenter keeps token use and latency low while maintaining high-quality video generation.

Demo Videos

Presentation examples

Example outputs in English and Chinese generated from scientific papers.

Benchmark

mPreBench for Effective Information Transfer

mPreBench is an expert-curated multilingual benchmark for evaluating whether generated presentation videos preserve the scientific information needed by viewers.

The benchmark covers 40 academic papers: 20 English papers from NeurIPS 2025 and ACL 2025, and 20 Chinese papers from Chinese Journal of Computers. Each paper is paired with expert-written multiple-choice questions translated into English, Chinese, German, Japanese, and Arabic.

40 academic papers
1,600 multilingual questions
5 evaluation languages
Benchmark
Topic distribution of the mPreBench dataset
mPreBench spans vision, NLP, graph, security, reinforcement learning, biomedical, and systems topics.

Results

Quality gains with practical efficiency

Ablations show the contribution of the Interpreter, layout inspection, and Reviewer. Efficiency analysis shows low latency and token consumption compared with baselines.

Ablation results for mPresenter
Each agentic component contributes to stronger QA accuracy and more wins over ablated variants.
Latency and token consumption comparison
mPresenter balances video quality, latency, and API token consumption for scalable generation.

Resources

Code and benchmark

The implementation and mPreBench benchmark resources are available on GitHub.

Code Benchmark Demo Videos

Citation

Reference

@inproceedings{han2026mpresenter,
  title     = {mPresenter: An Agentic Framework for Generating Multilingual Presentation Videos from Scientific Papers},
  author    = {Han, Wenhan and Xiao, Xiao and Pechenizkiy, Mykola and Fang, Meng},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year      = {2026}
}