Narrative-Driven Paper-to-Slide Generation via ArcDeck

Conference: ECCV2026
Paper: ECCV paper page
Area: Multi-Agent Systems
Keywords: slide generation, discourse trees, global commitment, narrative planning, multi-agent collaboration

The authors are Tarik Can Ozden, Sachidanand VS, Furkan Horoz, Ozgur Kara, Junho Kim, and James M. Rehg, affiliated with the University of Illinois Urbana-Champaign. This note uses the conference PDF; the cache does not establish an arXiv identifier or code URL for this paper, so neither is guessed here.

TL;DR

ArcDeck preserves argumentative dependencies through discourse trees, constrains multi-agent outline revision with a global commitment, and then constructs editable slides; with GPT-4o generation and GPT-5 judging on ArcBench, its Narrative Flow score is 43.25 versus SlideGen's 30.75, although overall quality remains below author-prepared slides.

Background & Motivation

Turning a paper into a talk is not simply compressing each section. Readers can revisit definitions, figures, and earlier arguments, whereas an audience receives information in slide order; showing the method before establishing its problem creates an unnecessary comprehension burden even when every slide is accurate. One-shot HTML generation can produce shallow summaries, while section-wise generation can lose dependencies across slides. Layout refinement cannot repair these narrative errors.

Systems such as PPTAgent, Paper2Poster, and SlideGen improve generation through template editing, visual feedback, or multi-agent collaboration, but producing a paper-level outline remains a difficult planning problem. Section headings indicate where content belongs without explaining whether a paragraph supplies context, supports a claim, or introduces a parallel contribution. ArcDeck makes that missing structure explicit: discourse trees represent local relationships, a separate document specifies global choices for the audience and duration, and the outline is then reviewed against both.

Core Idea: before generating layouts, make both "which material must be explained together" and "what the audience should remember" persistent constraints, then enforce them through critique, judgment, and revision of the slide narrative.

Method

Overall Architecture

The input consists of a paper PDF plus the target audience and presentation duration. Docling and Marker convert the main text into Markdown while excluding the bibliography and appendix; figures and tables enter an asset dictionary with captions and dimensions, and a separate citation map links short-form citations to numbered references.

The structured text feeds the Discourse Parser and Commitment Builder, which jointly guide the Narrative Refinement Loop. Its output is still a JSON outline rather than rendered pages. Deck Construction and Refinement then produces an editable .pptx through python-pptx. The seven agent roles in Figure 2 can be understood through the following four designs.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    Input["PDF, audience,<br/>and duration"] --> Prep["Text, assets,<br/>and citations"]
    Prep --> Parser["Discourse Parser"]
    Prep --> Commit["Commitment Builder"]
    Parser --> Loop["Narrative Refinement Loop"]
    Commit --> Loop
    Loop --> Render["Deck Construction<br/>and Refinement"]
    Prep --> Render
    Commit --> Render
    Render --> Output["Editable slide deck"]

Key Designs

1. Discourse Parser: group content by argumentative relationships

Inspired by Rhetorical Structure Theory (RST), the parser treats paragraphs as elementary discourse units (EDUs) and builds a binary tree within each paper section. Instead of independently scoring sentence importance, it assigns relationships to adjacent units and recursively combines them into larger spans. Relations near the root cover broader structure, while those near the leaves describe finer explanatory progression. The tree is serialized into JSON for the planner.

Table 1 distinguishes nucleus-satellite relations from multinuclear relations. The former connect a central claim with supporting material through Elaboration, Explanation, Context, Purpose, Evaluation, or Organization. The latter include Joint for equally central parallel units and Same-unit for a semantic unit split across boundaries. The distinction matters: an experimental detail supporting a methodological claim should not automatically be presented as an independent contribution.

The tree therefore guides grouping and ordering rather than prescribing one slide per node. The planner still decides which paragraphs belong together and records the rationale. The paper describes section-level trees connected through global planning, not a newly trained parser that establishes every rhetorical relationship across an entire document.

2. Commitment Builder: give every slide a shared presentation objective

Correct local explanations do not guarantee appropriate choices for the entire talk. The builder reads the Markdown, target audience, and duration to produce a Global Commitment comprising a snapshot, core content, talk contract, narrative spine, and light section plan. Core content establishes the thesis and takeaways; the talk contract specifies assumed knowledge, objectives, length, and presentation requirements; the section plan assigns purposes and priorities.

This document is a shared specification for planning and review, not another summary. It conditions both the refinement loop and deck construction, preserving the talk-level objective when individual slides are generated. Figure 4 illustrates a 20-minute talk for researchers with a target of 12–16 slides and explicit must-include and must-avoid content. This is an example configuration, not the fixed duration or slide count for every experiment.

Discourse trees and global commitments address different problems. The former preserve supporting relationships in the source; the latter determine which relationships deserve attention within the available time. Without the commitment, a deck can contain only source-grounded material yet order it as proposed method, existing methods, and task challenges. Figure 8 illustrates this type of narrative inversion.

3. Narrative Refinement Loop: separate identifying problems from authorizing revisions

The planner first groups paragraphs using the discourse trees and Global Commitment, producing a JSON outline with titles, content paragraphs, and grouping rationales. The critic then checks five dimensions against the commitment: objective alignment, global narrative flow, section balance, slide-level coherence, and redundancy or omissions. Reviewing the content plan makes missing prerequisites visible before they become polished but confusing slides.

Criticism is not executed unconditionally. The judge decides whether the outline is ready, explains its decision, assigns high, medium, or low severity to must-fix issues, and supplies actionable guidance. The reviser incorporates this feedback and returns the updated outline for evaluation. Processing stops when the judge approves or after at most 3 refinement cycles. The latter is a budget limit, not proof that every issue has been resolved.

Separating critique from judgment gives revision an explicit entry point and stopping rule instead of repeatedly asking the generator for an unconstrained rewrite. The object being improved is conceptual progression and coverage across slides. The final visual refinement stage cannot substitute for this loop and should not be confused with it.

4. Deck Construction and Refinement: realize the planned argument as visual pages

The constructor combines the outline, asset dictionary, and Global Commitment. It matches figures and tables using their captions, then selects among 14 layout templates according to text volume, asset count, dimensions, and aspect ratios. It chooses bullet hierarchies or paragraph narration based on information density, highlights key topics, and records short citations for later footnotes. The draft JSON contains global metadata and per-slide titles, text, visuals, and references.

The aesthetics agent receives asset-matching information, the outline, and the draft. It adds relevant figures where visual grounding is insufficient; enriches sparse content, condenses dense text, or merges underused layouts; derives a consistent theme color from frequently occurring colors in the figures; and bolds important terminology. python-pptx then performs rendering instead of requiring the model to emit a complete binary presentation.

This stage primarily reuses the paper's assets to express an already planned narrative rather than inventing arbitrary illustrations. Asset provenance and citations can be retained, but PDF extraction quality and the template library remain constraints. A coherent narrative does not automatically imply an optimal layout on every page.

A Worked Example

Consider the camera-pose research presentation illustrated in Figure 7. The intended progression is camera-pose prediction challenges, domain background, sparse-view pose estimation challenges, task introduction, and the proposed method. The parser preserves relationships between background and methodological explanations, the commitment establishes the central message, and the planner uses both to organize slide order.

If the draft introduces the task repeatedly or presents the method before necessary background, the critic identifies repetition and missing prerequisites. The judge selects the required corrections, and the reviser reorders or merges material. The constructor then assigns figures and layouts to method slides, followed by density refinement. This walkthrough interprets the paper's qualitative example; it is not an experimental trajectory with reported per-cycle scores.

Loss & Training

ArcDeck is an inference-time workflow using existing large language models and vision-language models (VLMs). The paper explicitly emphasizes RST parsing without fine-tuning and introduces no training loss or gradient-based optimization. Its main controls are structured prompt constraints, agent responsibilities, and at most 3 revision cycles. The revision loop should not be described as reinforcement learning training.

Key Experimental Results

Main Results

ArcBench selects 100 paper-slide pairs from 994 candidates collected across six CV/ML venues during 2022–2025, using this subset for all subsequent experiments. Papers must be oral presentations and contain at least 3 figures and 3 tables; the reference slides are author-prepared. These filters ensure demanding technical and visual material but do not constitute a random sample of academic presentations.

ArcDeck and four baselines use GPT-5, GPT-4o, and Qwen3-VL-32B generation backbones under the same source-paper and evaluation protocol. Generated slides use a 13.33 × 7.5-inch canvas and fixed themes. GPT-5 and Qwen3-VL serve as evaluators and must be distinguished from the generation backbone.

The quiz protocol in Table 2 lists 25 multiple-choice questions for each of Story, Visuals, Hard, and Depth. Text categories use extracted slide text, while Visuals uses slide images. Section 4.2 is somewhat ambiguous about question counts; this note follows the category-specific specification in Table 2. Accuracy is the fraction answered correctly. VLM-as-Judge evaluates text quality, narrative flow, layout, and visual theme with a 10-item checklist per dimension, reported on a 1–100 scale in the text. Quiz and A/B evaluations are repeated 11 times with randomized order, using averages or majority votes.

The following excerpt from Table 3 fixes GPT-4o generation and GPT-5 closed-source evaluation, avoiding comparisons that mix generation capability or evaluator preference. Hard and Depth are quiz accuracy percentages; NF and VL are judge scores.

Method	Hard ↑	Depth ↑	Narrative Flow NF ↑	Visual Layout VL ↑
HTML	11.84	14.16	14.00	24.00
Paper2Poster	17.60	20.24	28.12	60.25
PPTAgent	15.12	18.56	26.75	59.12
SlideGen	17.84	21.36	30.75	84.75
ArcDeck	40.08	47.20	43.25	91.00

Under this fixed condition, ArcDeck exceeds SlideGen by 22.24 percentage points on Hard, 25.84 percentage points on Depth, and 12.50 points on NF. Structural planning can improve information organization with the weaker generation backbone, but the same gain should not be assumed for every backbone.

Ablation Study

Table 5 reports the following results directly. "Full-method win rate" means ArcDeck's preference rate against the ablated version, not the ablated version's win rate. The main text does not clearly specify this table's generation backbone, evaluator, or NF scale conversion. Its original values are retained and should not be pooled with the NF values from Table 3.

Config	NF ↑	Full-method win rate (%)	Missing capability
Without Discourse Parser	7.50	89.1	Relation-guided content grouping
Without Global Commitment	7.52	94.5	Shared presentation objective and ordering constraints
Without Narrative Refinement Loop	8.68	61.8	Iterative outline review
ArcDeck	9.70	Not applicable	Full system

Removing either structural prior causes a larger decrease than removing the revision loop. However, individual removals establish the components' roles in this pipeline, not the optimal agent count or the uniquely best design under an equal budget.

Table 7 reports human evaluation with 25 undergraduate or MS/PhD students. Participants selected their most familiar area among three topics and ranked the three systems' decks for 5 oral papers in that area. First, second, and third places receive 3, 2, and 1 points respectively. These values are mean ranking points, not accuracy.

Method	Vision-Language / Multimodal	Generative Models	Computer Vision Core	Overall
PPTAgent	1.03	1.13	1.20	1.10
SlideGen	2.42	2.23	2.14	2.30
ArcDeck	2.55	2.63	2.66	2.60

Key Findings

• Table 4a does not show universal superiority: with Qwen3 generation and GPT-5 judging, the narrative preference rate against SlideGen is only 48.5%. With GPT-5 generation and GPT-5 judging, it is 50.6%, also close to parity.
• All six ArcDeck preference rates against author-prepared slides in Table 4b remain below 50%, with a maximum of 48.1%. Progress over automated systems is not evidence of surpassing the human reference.
• Figure 9 reports 128.5K tokens for outline generation and 83.3K tokens for slide generation. This is a substantial multi-call workflow; improved quality alone does not establish a cost advantage.

Highlights & Insights

• The reusable insight is to constrain two granularities: discourse relationships preserve local explanations, while the Global Commitment aligns the whole talk with one objective. For long reports or teaching materials, this is more inspectable than merely prompting the model to remain coherent.
• Revising the outline before visual composition localizes errors at the content-structure level. Moving a concept does not yet require rearranging every figure, giving content planning and page realization distinct responsibilities.
• Author-prepared references expose the blind spot of evaluating only relative baseline rankings. Winning among automated systems can still leave a substantial gap to presentation quality suitable for direct delivery.

Limitations & Future Work

• The authors explicitly leave expansion to other disciplines for future work. AI venues, oral papers, and figure-rich articles do not establish generalization to medical talks, purely theoretical work, or nontechnical business presentations.
• This note's evaluation assessment: VLMs participate in both generation and judging. Open- and closed-source judges help inspect bias, but their scores still differ markedly; rankings from 25 participants also do not directly measure audience learning.
• The main text does not provide sufficient Table 5 evaluation configuration, statistical uncertainty, or equal-token-budget baselines. These are needed to better separate structural design, additional model calls, and backbone capability.
• Models generate both discourse trees and commitments, and the main text does not report relation-label accuracy. An initial misunderstanding may become a constraint that all later agents consistently follow; paragraph provenance checks and factual verification would be useful additions.
• Excluding appendices, restricting layouts to 14 templates, and allowing at most 3 revision cycles constrain coverage and design freedom. Longer or more complex talks need explicit quality, latency, and cost trade-off measurements.

Related Work & Insights

• Compared with SlideGen: both use collaborative multimodal agents. ArcDeck's distinction is not the first use of multiple agents, but RST relationships and a Global Commitment before outline generation, followed by dedicated narrative revision. Near-parity and below-parity results in Table 4a also rule out uniformly dominant performance.
• Compared with PPTAgent: PPTAgent emphasizes reference-template editing, whereas ArcDeck reconstructs argumentative relationships before instantiating templates. The former addresses page reuse and editing, while the latter more directly targets dependencies within technical content.
• Compared with Paper2Poster: the paper adapts a poster-generation system into a slide baseline. Results therefore describe that adaptation protocol rather than invalidating visual-feedback generation in general.
• Compared with RST-based text generation: the contribution mainly demonstrates the value of established rhetorical modeling for presentation planning, rather than introducing a new linguistic relation taxonomy. The transferable question is how explicit relationships constrain long-document generation, not simply how many agent roles to add.

Rating

• Novelty: 4/5. Discourse relationships, global talk constraints, and independent outline revision form a focused contribution to narrative planning.
• Experimental Thoroughness: 4/5. Three generation backbones, two model-judge families, ablations, and human rankings provide useful coverage, with gaps in cost control and ablation configuration.
• Writing Quality: 4/5. Module inputs and outputs are clear, but some table scales and narrative summaries require careful checking.
• Value: 4/5. Useful for drafting technical presentations and establishing author-prepared references, but not a substitute for human review before delivery.