Narrative-Driven Paper-to-Slide Generation via ArcDeck
| Authors | Tarik Can Ozden et al. |
| Year | 2026 |
| HF Upvotes | 6 |
| arXiv | 2604.11969 |
| Download | |
| HF Page | View on Hugging Face |
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.
Engineering Breakdown
Plain English
ArcDeck is a system that converts academic papers into presentation slides by explicitly modeling the logical structure and narrative flow of the source document rather than just summarizing raw text. The approach first parses the paper into a discourse tree and creates a global commitment document that captures the high-level research intent, then uses multiple specialized AI agents that iteratively critique and refine a presentation outline before generating the final visual layouts. The authors also introduced ArcBench, a new benchmark of paper-slide pairs, to evaluate their method. Their experimental results demonstrate that explicitly modeling discourse structure produces better presentations than methods that skip this structural analysis step.
Core Technical Contribution
The key innovation is formulating paper-to-slide generation as a structured narrative reconstruction task rather than a direct text summarization problem. ArcDeck explicitly constructs a discourse tree during parsing to capture how the paper's arguments and claims flow logically, then uses this structural prior to guide a multi-agent refinement loop. This stands in contrast to prior approaches that treat slides as compressed summaries of raw text without modeling the underlying argumentative structure. The framework combines discourse analysis, structured document modeling (via commitment documents), and iterative multi-agent refinement into a cohesive pipeline that preserves research intent.
How It Works
The pipeline operates in three main stages: First, the input paper is parsed to construct a discourse tree that models logical relationships between sections and claims, and a global commitment document is generated to capture the paper's high-level research goals and narrative arc. Second, specialized agents work iteratively—one agent proposes a presentation outline based on the discourse tree, another critiques it against the commitment document and the original paper, and a third refines the outline based on feedback, repeating until convergence. Third, the final outline is rendered into visual layouts and slide designs with appropriate typography, spacing, and visual hierarchy. The discourse tree acts as a structural prior that constrains the agent-based refinement process, ensuring the presentation maintains the original paper's logical flow rather than arbitrarily reordering content.
Production Impact
For teams building document-to-presentation automation tools, this approach would significantly improve output quality by capturing and preserving research intent—critical for academic and technical contexts where slide order and narrative coherence matter. In a production pipeline, you would add a discourse parsing stage (likely using a specialized LLM fine-tuned on academic papers) and implement a multi-agent loop with checkpointing to handle iterative refinement, increasing latency from seconds to potentially minutes depending on paper length. The multi-agent architecture adds orchestration complexity—you need agent state management, consensus logic, and convergence criteria—but the improvement in semantic fidelity would justify this overhead for high-stakes use cases like conference presentations or grant proposals. The commitment document mechanism could also be exposed as a user-facing control: researchers could explicitly state "this paper's main contribution is X" and constrain the system to ensure slides prominently feature X, making the system more controllable and trustworthy.
Limitations and When Not to Use This
The paper's abstract doesn't specify computational requirements, but the multi-agent iterative refinement process is inherently expensive—likely requiring multiple forward passes through large language models and potentially heavy discourse parsing overhead that may not be acceptable for real-time or mobile use cases. The approach assumes the input paper is well-structured with clear logical flow; it may struggle with papers that have unconventional organization, dense mathematical notation without natural language scaffolding, or highly specialized terminology that the discourse parser hasn't seen during training. The system requires a fine-tuned discourse parser and likely domain-specific training data, so generalization to non-academic documents (business reports, technical specifications, marketing materials) is unclear. The introduced ArcBench benchmark, while valuable, is limited to academic papers, leaving open questions about performance on diverse document types and whether the approach scales to very long papers (100+ pages) where discourse trees become unwieldy.
Research Context
This work builds on a growing body of research in document understanding and summarization that moves beyond extractive or abstractive approaches toward structured, intent-preserving transformations. It extends prior work in discourse parsing and rhetorical structure theory (RST) by applying these frameworks to a practical generation task, demonstrating that explicit structural modeling improves downstream application performance. The multi-agent refinement pattern parallels recent trends in complex reasoning systems (chain-of-thought prompting, self-critique loops) and shows these techniques are effective for high-stakes document transformation tasks. The introduction of ArcBench fills a gap in public benchmarks for paper-to-slide generation, similar to how datasets like SQuAD and HuggingFace transformers have accelerated progress in other NLP tasks, and opens the door for future work on architectural improvements and cross-domain generalization.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::