# Narrative-Driven Paper-to-Slide Generation via ArcDeck

Source: [Narrative-Driven Paper-to-Slide Generation via ArcDeck](<https://www.sota2.com/research/paper/69df1b987d0172140ec7b92e>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-04-19T06:00:01.402Z

Authors: Tarik Can Ozden, Sachidanand VS, Furkan Horoz, Ozgur Kara, Junho Kim, James Matthew Rehg

Published: 2026-04-13T19:03:03.000Z

## Abstract

We introduce ArcDeck, a multi-agent framework that formulates paper-to-slide generation as a structured narrative reconstruction task. Unlike existing methods that directly summarize raw text into slides, ArcDeck explicitly models the source paper's logical flow. It first parses the input to construct a discourse tree and establish a global commitment document, ensuring the high-level intent is preserved. These structural priors then guide an iterative multi-agent refinement process, where specialized agents iteratively critique and revise the presentation outline before rendering the final visual layouts and designs. To evaluate our approach, we also introduce ArcBench, a newly curated benchmark of academic paper-slide pairs. Experimental results demonstrate that explicit discourse modeling, combined with role-specific agent coordination, significantly improves the narrative flow and logical coherence of the generated presentations.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [VLM Pairwise Preference on ArcBench](<https://www.sota2.com/research/sota/vlm-pairwise-preference-on-arcbench.md>): VLM Pairwise Preference; ArcBench; Win Rate: 100.

- [Presentation Generation on Presentation Generation Evaluation Set](<https://www.sota2.com/research/sota/presentation-generation-on-presentation-generation-evaluation-set.md>): Presentation Generation; Presentation Generation Evaluation Set; ROUGE-L: 156.

- [VLM-based Q/A Quiz on ArcBench VLM-based Q/A Quiz Open](<https://www.sota2.com/research/sota/vlm-based-q-a-quiz-on-arcbench-vlm-based-q-a-quiz-open.md>): VLM-based Q/A Quiz; ArcBench VLM-based Q/A Quiz Open; Accuracy (Story): 95.69.

- [Slide Deck Generation Quality Evaluation on Five Oral Papers Subset (V-L/MM, Gen. Models, CV Core) (test)](<https://www.sota2.com/research/sota/slide-deck-generation-quality-evaluation-on-five-oral-papers-subset-v-l-mm-gen-models-cv-core-test.md>): Slide Deck Generation Quality Evaluation; Five oral papers from V-L/MM, Generative Models, and CV Core (test); Score (V-L/MM): 2.55.

- [VLM-as-Judge Evaluation on ArcBench VLM-as-Judge Open](<https://www.sota2.com/research/sota/vlm-as-judge-evaluation-on-arcbench-vlm-as-judge-open.md>): VLM-as-Judge Evaluation; ArcBench VLM-as-Judge Open; TQ: 80.06.

- [VLM-as-Judge Evaluation on ArcBench VLM-as-Judge Closed](<https://www.sota2.com/research/sota/vlm-as-judge-evaluation-on-arcbench-vlm-as-judge-closed.md>): VLM-as-Judge Evaluation; ArcBench VLM-as-Judge Closed; TQ: 53.62.

- [VLM-based Q/A Quiz on ArcBench VLM-based Q/A Quiz Closed](<https://www.sota2.com/research/sota/vlm-based-q-a-quiz-on-arcbench-vlm-based-q-a-quiz-closed.md>): VLM-based Q/A Quiz; ArcBench VLM-based Q/A Quiz Closed; Story Accuracy: 96.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2604.11969>)

[PDF](<https://arxiv.org/pdf/2604.11969.pdf>)

[Implementation](<https://github.com/RehgLab/ArcDeck>)