Source-linked AI summary
PPTAgent: Generating and Evaluating Presentations Beyond Text-to-Slides
Hao Zheng, Xinyan Guan, Hao Kong, Jia Zheng, Weixiang Zhou, Hongyu Lin, Yaojie Lu, Ben He, Xianpei Han, Le Sun
TL;DR
Automatic presentation generation often emphasizes textual content while neglecting visual appeal and structural coherence. PPTAgent addresses this gap with a two-stage, reference-guided editing workflow, while PPTEval evaluates Content, Design, and Coherence; PPTAgent achieves an average score of 3.67 across the three dimensions.
Problem
Existing presentation-generation methods focus primarily on textual content and neglect the visual and structural aspects needed for well-rounded presentations.
Method
PPTAgent analyzes reference presentations for slide functional types and content schemas, then drafts an outline and iteratively edits selected reference slides using executable actions.
Results
PPTAgent achieves an average PPTEval score of 3.67 across Content, Design, and Coherence and significantly improves over baseline methods across all three dimensions.
Takeaways & Limitations
The paper presents reference-guided editing and multidimensional evaluation as a new paradigm for automatic presentation generation under unsupervised conditions.
Takeaways & Limitations
Generated presentation quality remains influenced by the input reference, and occasional failures and overlapping elements limit reliability and readability.
Abstract
from arXiv · showhide
Automatically generating presentations from documents is a challenging task that requires accommodating content quality, visual appeal, and structural coherence. Existing methods primarily focus on improving and evaluating the content quality in isolation, overlooking visual appeal and structural coherence, which limits their practical applicability. To address these limitations, we propose PPTAgent, which comprehensively improves presentation generation through a two-stage, edit-based approach inspired by human workflows. PPTAgent first analyzes reference presentations to extract slide-level functional types and content schemas, then drafts an outline and iteratively generates editing actions based on selected reference slides to create new slides. To comprehensively evaluate the quality of generated presentations, we further introduce PPTEval, an evaluation framework that assesses presentations across three dimensions: Content, Design, and Coherence. Results demonstrate that PPTAgent significantly outperforms existing automatic presentation generation methods across all three dimensions.
1 Introduction
PPTAgent addresses limitations in text-focused presentation generation by modeling slide structure and using an edit-based workflow guided by reference presentations. PPTEval evaluates generated presentations across Content, Design, and Coherence, with PPTAgent achieving an average score of 3.67.
- Motivation: Existing text-to-slides approaches treat presentation generation primarily as abstractive summarization, neglecting visual design and producing text-heavy, monotonous presentations.These approaches focus mainly on textual content despite presentations’ visual-centric nature.
- PPTAgent: PPTAgent reframes slide generation as selecting reference slides and editing them step by step, following a workflow inspired by human presentation creation.The approach decomposes generation into reference selection and iterative editing rather than creating complex slides from scratch in one pass.
- PPTAgent: PPTAgent analyzes reference presentations to extract slide functional types and content schemas, supporting reference selection and subsequent slide generation.The analysis includes clustering reference slides into categories such as opening and bullet-point slides.
- PPTAgent: PPTAgent uses edit APIs with HTML-rendered slide representations and self-correction to refine editing actions from execution feedback.This design addresses the verbosity of PowerPoint XML and enables iterative refinement when actions fail.
- Evaluation: PPTEval assesses presentation quality across Content, Design, and Coherence, and PPTAgent achieves an average score of 3.67 across these dimensions.Human evaluations are reported to validate PPTEval’s reliability and effectiveness.
- Resources: The authors release PPTAgent, PPTEval, and the Zenodo10K presentation dataset to support future research.The releases include codebases for both the generation framework and evaluation framework.
2 PPTAGENT
PPTAgent formulates presentation generation as a two-stage editing task: it analyzes reference slides, then selects and iteratively edits them to generate target presentations. Its APIs and feedback loop are designed to preserve layouts while improving editing robustness.
- Framework Overview: PPTAgent consists of Stage I reference-presentation analysis and Stage II reference-guided target-slide generation through iterative editing.The two stages provide structural information before slides are generated from selected references.
- Task Formulation: Conventional generation maps input content directly to slide elements with manually specified styling attributes, whereas PPTAgent generates executable actions to edit a reference slide.The edit-based formulation is intended to preserve existing layouts and styles.
- Stage I: Presentation Analysis: Stage I clusters slides by structural and layout characteristics, distinguishing structural slides from content slides.Structural slides are grouped using textual features, while content slides are grouped through image-based hierarchical clustering.
- Stage I: Presentation Analysis: Schema extraction represents each slide element by its category, description, and content to provide a structured account of slide organization.The extracted schemas facilitate subsequent slide generation.
- Stage II: Presentation Generation: PPTAgent generates an outline whose entries pair each new slide with a reference slide and relevant document content.Reference selection uses slide-level functional descriptions, while relevant content is identified from the input document.
- Stage II: Presentation Generation: Slides are generated iteratively by editing reference elements through APIs that support editing, removing, and duplicating elements.Reference slides are rendered as HTML to make modifications more precise and intuitive than direct XML editing.
- Stage II: Presentation Generation: A self-correction mechanism executes editing actions in a REPL and uses execution feedback to refine invalid actions until a valid slide is generated.The loop lets the language model adjust actions based on intermediate execution results.
3 PPTEVAL
PPTEval is a reference-free framework for assessing presentation quality with numeric scores and rationales. It evaluates Content and Design at the slide level and Coherence across the full presentation.
- Framework: PPTEval addresses the absence of comprehensive reference-free presentation evaluation by providing numeric scores and detailed rationales.Its scores use a 1-to-5 scale.
- Dimensions: Content and Design are assessed at the slide level, while Coherence is evaluated across the entire presentation.This separates local slide quality from whole-presentation organization.
- Dimensions: PPTEval evaluates presentations across three dimensions: Content, Design, and Coherence.The framework is grounded in established presentation design principles.
- Scoring: PPTEval uses a five-point scoring scale for its evaluation criteria.The scoring criteria are summarized in Table 1.
4 Experiment
Experiments evaluate Zenodo10K-based presentation generation with standard metrics and PPTEVAL. PPTAGENT improves presentation quality across Content, Design, and Coherence and shows robust generation performance.
- 4.1 Dataset: Zenodo10K addresses semantic-information loss and limited domain diversity in existing presentation datasets.The experiments sample presentations and documents across five domains.
- 4.3 Evaluation Metrics: PPTEVAL evaluates Content and Design at the slide level and Coherence at the presentation level.It uses GPT-4o as the judging model and averages Content and Design scores across slides.
- 4.5 Overall Result: PPTAGENT significantly outperforms DocPres and KCTV across all three PPTEVAL dimensions.Against DocPres, it improves Design from 2.37 to 3.34 (+40.9%) and Content from 2.98 to 3.34 (+12.1%); against KCTV, it improves Design from 2.95 to 3.34 (+13.2%) and Content from 2.55 to 3.28 (+28.6%).
- 4.5 Overall Result: 4.48 vs. 3.57 and 3.28: PPTAGENT improves Coherence by 25.5% over DocPres and 36.6% over KCTV.The authors attribute this improvement to comprehensive analysis of the structural role of slides.
- 4.5 Overall Result: 97.8% vs. 88.0%: PPTAGENT achieves a higher Success Rate than KCTV, while both tested PPTAGENT configurations reach at least 95%.Success requires all slides to be generated without execution errors after self-correction for PPTAGENT.
- 4.6 Evaluation Analysis: PPTEVAL exposes evaluation differences that PPL and ROUGE-L do not consistently capture.KCTV has ROUGE-L 16.76 but Content 2.55, whereas PPTAGENT has ROUGE-L 14.25 and Content 3.28; PPTEVAL also supports reference-free Design assessment and holistic Coherence evaluation.
5 Analysis
Analysis examines ablations, score distributions, self-correction, and evaluator agreement. The results connect structural analysis and HTML-based editing with generation quality, while PPTEVAL aligns with human preferences.
- 5.1 Ablation Study: Removing CodeRender lowers Success Rate from 95.0% to 74.6%, indicating that HTML-based representation reduces interaction complexity.The ablation replaces the slide representation used by PPTAGENT.
- 5.1 Ablation Study: Removing Outline or Structure lowers Coherence from 4.48 to 3.36 or 3.45, while removing Schema lowers Success Rate from 95.0% to 78.8%.These ablations test reference selection, structural-slide handling, and content-schema guidance.
- 5.2 Case Study: More than 80% of PPTAGENT presentations score at least 3 in Content and Design, and more than 80% score above 4 in Coherence.Baseline scores are concentrated mainly at levels 2 and 3 for Content and Design.
- 5.3 Self-Correction Analysis: GPT-4o has stronger self-correction than Qwen2.5, while Qwen2.5 makes fewer first-generation errors; all three models correct more than half of errors.Qwen2-VL encounters errors more frequently and has poorer self-correction capabilities in the reported investigation.
- 5.5 Human Preference Agreement: 0.71: PPTEVAL’s average Pearson correlation with human ratings exceeds that of other evaluation methods.The reported correlations across Coherence, Content, and Design are statistically significant.
- 5.6 Existing Metrics: PPL, ROUGE-L, and FID show weak or inconsistent relationships with PPTEVAL’s presentation-quality dimensions.The analysis specifically notes that PPL measures text fluency but performs poorly on slide Content because of presentation fragmentation.
6 Related Works
Related work spans rule-based and template-based slide generation, LLM agents, and LLM-as-a-judge evaluation. PPTAGENT is positioned against methods that limit visual coverage, scalability, or multidimensional assessment.
- Automated Presentation Generation: Rule-based methods emphasize textual content but neglect presentations’ visual-centric nature, producing less engaging outputs.The section cites Mondal et al. and DocPres as examples.
- Automated Presentation Generation: Template-based methods can produce visually appealing slides but depend on extensive manual template annotation, limiting scalability and flexibility.KCTV is identified as a template-based method.
- LLM Agent: LLM-agent research explores executable actions for task automation, including presentation-related API integration.These studies motivate action-based interaction with presentation software.
- LLM as a Judge: LLM-as-a-judge research supports using multimodal models for evaluation, but prior work did not assess presentation quality across the full set of dimensions addressed here.The cited literature includes multidimensional and single-slide evaluation efforts.
7 Conclusion
PPTAGENT reframes presentation generation as a two-stage editing task using LLMs to understand and generate code, while PPTEVAL quantitatively assesses presentation quality. Experiments across multiple domains demonstrate the method’s superiority and support a new unsupervised slide-generation paradigm.
- PPTAGENT conceptualizes presentation generation as a two-stage editing task completed through LLM-based understanding and code generation.
- PPTEVAL provides quantitative metrics for assessing presentation quality across the proposed evaluation framework.
- Experiments across data from multiple domains demonstrate PPTAGENT’s superiority over existing approaches.
- The research offers a new paradigm for generating slides under unsupervised conditions and insights for future presentation-generation work.
Limitations
PPTAGENT remains constrained by occasional generation failures, dependence on reference-presentation quality, and incomplete use of visual information for slide-design refinement.
- Despite a success rate above 95% on the dataset, PPTAGENT occasionally fails to generate presentations, limiting reliability.
- Generated-presentation quality remains influenced by the input reference presentation, which can produce suboptimal outputs.
- PPTAGENT does not fully use visual information to refine slide design, resulting in occasional flaws such as overlapping elements.
- Future work should improve robustness, reduce reference dependency, and incorporate visual information more effectively.
Ethical Considerations
Zenodo10K construction followed licensing requirements and compensated annotation personnel above applicable minimum-wage rates.
- Artifacts lacking permission for modification or commercial use were filtered out to comply with intellectual-property rights.
- Annotation personnel were compensated at rates exceeding the minimum wage in their respective cities.
A Data Preprocessing
The preprocessing pipeline bounded source-document and presentation sizes, extracted textual and visual content, generated image captions, and removed near-duplicate images and slides using embedding similarity thresholds.
- Presentations were limited to 12–64 pages and documents to 2,048–20,480 characters to maintain a reasonable cost.
- Textual and visual content was extracted from source documents, with image captions generated to support relevant image selection through textual descriptions.
- Duplicate images were removed when their image-embedding cosine similarity exceeded 0.85.
- Individual slides were removed when their text-embedding cosine similarity exceeded 0.8 compared with the preprocessed set.
B Details of PPTEVAL
PPTEVAL evaluates presentations across Content, Design, and Coherence, using slide images, multimodal models, and structured scoring criteria. Its supporting pipeline analyzes structural roles, clusters content slides by layout similarity, and enables edit-based presentation generation through executable actions.
- Human evaluation: Four graduate students evaluated 250 presentations, including real-world presentations and outputs from a baseline and PPTAGENT.Evaluators scored converted slide images individually according to PPTEVAL criteria.
- Human evaluation: 0.59 average Fleiss’ Kappa indicated satisfactory inter-rater agreement across Content, Design, and Coherence.The dimension-specific values were 0.61, 0.61, and 0.54, respectively.
- Evaluation dimensions: PPTEVAL evaluates Content, Design, and Coherence using converted slide images and five-point scoring criteria.Content covers information amount, textual clarity, and visual support; Design covers color schemes, visual elements, and overall design; Coherence covers logical structure and contextual information.
- Reference analysis: Slides are grouped by layout similarity after text and images are neutralized, using ViT-image embeddings and a similarity threshold θ of 0.65.Hierarchical clustering focuses on layout patterns rather than specific content, after which multimodal models identify patterns within clusters.
- Presentation generation: PPTAGENT edits selected reference slides through executable actions instead of creating slides from scratch, preserving their layouts and styles.The generation prompts cover outline creation, slide content, and editing actions, while a code generator translates commands into API calls using HTML structure and parent-child relationships.