AI Powered Presentation Platforms

• AI-PPPs are integrated systems that harness LLMs, multimodal models, retrieval, and speech technologies to automate drafting, editing, and delivering presentations.
• They rely on structured content extraction from documents and notebooks, using methods like hierarchical parsing and semantic matching to generate grounded, executable slide decks.
• These platforms extend traditional slide generation by incorporating live adaptation, personalized delivery, and coaching feedback to enhance presentation effectiveness.

AI Powered Presentation Platforms (AI-PPPs) are presented in recent arXiv literature as systems that use LLMs, multimodal models, retrieval, code generation, dialogue, and speech technologies to support the presentation lifecycle: transforming documents or notebooks into draft decks, designing slides from natural-language intent, adapting visuals during live discussion, synthesizing narrated videos, supporting rehearsal, and delivering rubric-based or coaching-oriented feedback (Fu et al., 2021, Wang et al., 2024, Ge et al., 1 Jan 2025, Takahira et al., 16 Oct 2025, Shi et al., 5 Jul 2025, Becerra et al., 6 May 2026). Taken together, these works show that presentation support is no longer treated as a narrow text-to-slides problem; content planning, audience adaptation, delivery, and evaluation increasingly appear as first-class presentation tasks (Yang et al., 1 Apr 2026).

1. Scope and historical development

Early work framed presentation generation primarily as multimodal summarization and layout prediction. DOC2PPT defined document-to-slide generation over sections, sentences, figures, and layouts, and released 5,873 paired documents and slide decks from CV, NLP, and ML venues (Fu et al., 2021). In parallel, the 2021 presentation creator with customized audio delivery targeted structured documents such as research papers, combining BERT-based extractive summarization with slide creation and customized narration via a Tacotron-inspired synthesizer and GAN-based vocoder (Mansoor et al., 2021).

By 2022, the scope broadened in two different directions. NB2Slides moved into notebook-native authoring for data science work, emphasizing audience-aware structuring, linked notebook-slide explanations, and human-AI co-authoring inside JupyterLab (Zheng et al., 2022). Pre-Avatar shifted attention from slide authoring to speaker-personalized presentation video generation, using a talking avatar derived from 1 frontal face photo and a 3-minute voice recording (Sun et al., 2022).

Subsequent work expanded along at least four axes. PPTC made multi-turn, multimodal PowerPoint task completion itself an evaluation target (Guo et al., 2023). Outline-first and agentic authoring systems such as OutlineSpark, PPTAgent, PreGenie, Auto-Slides, and SlideBot made planning, retrieval, verification, and executable intermediate representations explicit (Wang et al., 2024, Zheng et al., 7 Jan 2025, Xu et al., 27 May 2025, Yang et al., 14 Sep 2025, Xie et al., 12 Nov 2025). VisAider and DeepSlide introduced presentation-time adaptation and delivery-aware optimization (Takahira et al., 16 Oct 2025, Yang et al., 1 Apr 2026). AISSA and PresentCoach treated slides and rehearsals as assessable, revisable objects within educational feedback loops (Becerra et al., 6 May 2026, Chen et al., 19 Nov 2025). This chronology indicates that AI-PPPs now span authoring, live support, delivery, evaluation, and post-presentation transformation rather than a single generation task.

2. Source-grounded authoring from documents and notebooks

Document- and notebook-grounded platforms typically separate source understanding, relevance matching, and rendering. DOC2PPT uses a hierarchical recurrent sequence-to-sequence architecture with four modules—Document Reader, Progress Tracker, Object Placer, and Paraphraser—and jointly optimizes structure, object selection, text-figure matching, paraphrasing, and 4D bounding-box layout through

$$\mathcal{L}_{DOC2PPT} = \mathcal{L}_{structure} + \gamma \mathcal{L}_{content},$$

with $\gamma = 1$ (Fu et al., 2021). The system is explicitly a draft-generation assistant rather than a replacement for human presenters.

NB2Slides treats the notebook as technical evidence and the slide outline as rhetorical scaffold. It parses a notebook into a tree, embeds slide-section descriptions and notebook cells with a supervised SimCSE sentence embedding model, computes cosine similarity, and retrieves top-$k$ cells for each section, with $k=3$ for most queries and $k=1$ for some sections such as data source. It then uses a CodeTrans T5-based summarizer to convert code plus associated documentation into slide-ready text, while user-specified audience background and level of details affect both outline choice and sentence length (Zheng et al., 2022).

OutlineSpark reorders this workflow around ideation. As a JupyterLab plugin, it first summarizes each cell with up to five representative keywords, then lets the user draft a hierarchical outline, flattens that outline into context-rich “outline units,” maps each unit to up to five relevant notebook cells with semantic relevance scores from 0 to 1, and generates titles, bullet points, and visual elements with GPT-3.5-turbo-16k at temperature 0 (Wang et al., 2024). Auto-Slides generalizes the same planning emphasis to research papers: a Parser Agent based on Marker extracts figures, tables, and equations into JSON, a Planner Agent rewrites IMRaD structure into a PMRC narrative, Verification and Adjustment agents repair missing high-importance content, and a Generator Agent emits LaTeX Beamer code that can be modified through a ReAct-style Editor Agent (Yang et al., 14 Sep 2025).

These systems share a common premise: source artifacts are not treated as undifferentiated text. Instead, notebook cells, equations, figures, captions, tables, and section hierarchies are preserved as structured evidence so that slide generation can be grounded, inspectable, and revisable.

3. Agentic generation and executable slide representations

A major design shift in AI-PPPs is the move from direct slide synthesis to executable intermediate representations. PPTAgent first analyzes reference presentations to infer functional slide roles and content schemas, drafts an outline that specifies the reference slide to reuse, source-document section indices, and the target slide’s title and description, and then generates a new deck as a sequence of editing actions such as del_span, del_image, clone_paragraph, replace_span, and replace_image. The generation target is therefore an action sequence,

$\boldsymbol{A}=\sum_{i=1}^{m} a_i = f(C \mid R_j),$

rather than a slide directly (Zheng et al., 7 Jan 2025).

AutoPresent makes a similar argument in the natural-language-to-slide setting. Its SlidesBench benchmark contains 7k training and 585 testing examples derived from 310 slide decks across 10 domains, and it compares end-to-end image generation against program generation. The paper finds that programmatic methods produce higher-quality slides in user-interactable formats, and its SlidesLib abstraction reduces average program length from about 170 lines to about 13 lines through higher-level APIs such as add_title, add_text, add_bullet_points, add_image, set_background_color, search_image, and generate_image (Ge et al., 1 Jan 2025).

PreGenie adopts Slidev Markdown as the rendering substrate and organizes generation into Analysis and Initial Generation followed by Review and Re-generation. Its Text Summarizer, Image Captioner, Code Generator, Code Reviewer, and Page Reviewer share context so that syntax errors, formatting problems, overflow, cramped layouts, and text-image inconsistencies can be caught first at code level and then again after rendering (Xu et al., 27 May 2025). SlideBot applies a comparable decomposition in educational settings, with a Retriever, Summarizer, Moderator, Code Generator, and Enhancer. It grounds content via arXiv API retrieval or a biomedical textbook corpus ranked with BM25, generates LaTeX Beamer, and augments slides with figure macros and togglable pdfcomment annotations for instructor-facing refinement (Xie et al., 12 Nov 2025).

This family of systems treats presentations as executable structured artifacts rather than static images. A plausible implication is that editability, verification, and post-generation repair become architectural properties, not merely interface conveniences.

4. Live adaptation, delivery automation, and post-presentation transformation

AI-PPPs increasingly address presentation-time behavior instead of only precomputed decks. VisAider is a multi-stage LLM-based pipeline, orchestrated with LangChain and powered by GPT-4o, with five modules: Content analysis, Data selection, Candidate generation, Vega specification generation, and Candidate evaluation. It continuously analyzes the available dataset, active visualization, ongoing conversation, and audience profile to produce a ranked list of visualization aids that may differ in chart type, encoding, title, data range, dataset choice, and transformations. The current prototype generates eight ranked candidates in about 20–30 seconds, which the authors identify as too slow for live presentation flow (Takahira et al., 16 Oct 2025).

On the delivery side, PASS divides the workflow into Slide Generation and Slide Presentation. From a general Word document it uses an Image and Text Extractor, Title Generator, Content Extractor, Summarizer, Image Mapping module, Presenter Script Generator, and Audio Generation module; it produces up to 8–10 slides, generates per-slide scripts, and synthesizes narration with a Tacotron-2-based text-to-speech model implemented through SpeechBrain. On the SciDuet test dataset of 81 research papers, GPT-PASS reports 8.79 coherence, 8.34 redundancy, 9.75 relevance, and 8.96 average, while Qwen-PASS reports 8.65, 8.35, 9.68, and 8.89 (Aggarwal et al., 11 Jan 2025). PresentAgent likewise turns long-form documents into narrated presentation videos through document segmentation, slide planning, oral-style narration, MegaTTS3 speech synthesis, and FFmpeg composition, with PresentEval scoring content fidelity, visual clarity, and audience comprehension (Shi et al., 5 Jul 2025).

DeepSlide extends this delivery orientation by adding requirement elicitation, multiple logical-chain candidates with per-node time budgets, a lightweight content-tree retriever, attention augmentation such as figure zoom/focus, table-to-visualization conversion, text-to-diagram transformation, text keynote highlighting, auto layout, motion, and rehearsal support (Yang et al., 1 Apr 2026). Some platforms push further, beyond delivery into durable scholarly transformation. AI-Reporter uses Partitur pipelines, slide-video synchronization, contextual transcription, narrative synthesis, and academic contextualization to turn a presentation PDF and video recording into a publication-ready chapter in less than three minutes; in the reported case, a 16-minute presentation became 14 chronological storyboard blocks and then 13 thematically organized sections (Graßhoff, 8 Jul 2025). Speaker-personalized delivery remains a parallel thread: Pre-Avatar generates a full presentation video from slides, notes, 1 frontal face photo, and a 3-minute voice recording, while earlier voice-cloning work coupled document summarization and customized audio delivery from a short audio clip (Sun et al., 2022, Mansoor et al., 2021).

5. Coaching, formative assessment, and educational deployment

A distinct branch of AI-PPPs treats the slide deck as an object of assessment and revision. AISSA is a modular web application built with Python and Plotly Dash, organized into a Visualization Module, Processing and Analysis Module, Data Persistence Module, Extraction Module, and AI Module. Using python-pptx and OpenCV, it extracts word counts, font sizes, font families, reference-related features, edge density, and color dimensionality, then submits the teacher rubric and extracted features to ChatGPT 5.2 through a restrictive GePeTo-derived prompt that enforces a predefined JSON schema with 1–5 rubric scores, overall score and percentage, and bilingual qualitative feedback in Spanish and English. In a pilot with 46 undergraduate students, AISSA processed 90 presentations in about 1 to 3 minutes per submission at an average cost of $0.06–$0.07 USD and obtained a mean SUS of 83.38 (Becerra et al., 6 May 2026).

PresentCoach frames rehearsal as a closed loop of observation, practice, and reflection. Its Ideal Presentation Agent converts a .pptx deck into a model presentation video by rendering slides to high-resolution PNGs, generating a coherent English narration script of typically 60–100 words per slide with Qwen2.5-VL, cloning the user’s voice with CosyVoice2, and composing synchronized output with FFmpeg. Its Coach Agent then compares user audio against slide image, ideal narration script, and ideal audio using Gemini 2.5 Pro, returning encouragement followed by Observation–Impact–Suggestion feedback under 150 words. In a controlled between-subjects study with $N = 24$, the PresentCoach+PPT condition improved PRCS from 3.69 ± 0.81 to 5.03 ± 0.63 (+36.3%, $p = .016$), whereas PPT-only changed from 3.18 ± 0.65 to 3.69 ± 1.17 (+16.1%, $p = .219$) (Chen et al., 19 Nov 2025).

The JAM-based public-speaking platform adapts the “Just a Minute” game into a Flask-based local web app using gpt-4o-mini, tts-1, and NLTK. It measures performance with

$\gamma = 1$0

and

$\gamma = 1$1

but also reports that about 35% of word tokens were incorrectly recognized by speech transcription (Higham et al., 3 Oct 2025). In classroom deployment, AI-PPP outputs need not be perceptually separable from human-produced slides. In a Web Software Development course at Aalto University, coding-assistant workflows based on Cursor and Claude Code were judged by educators to be the most accurate, complete, and pedagogically sound among five tools, and students rated AI-generated slides and human-generated slides identically at mean 5.1 and median 5.0; Mann–Whitney tests showed no evidence of a difference in perceived quality $\gamma = 1$2 and no reliable ability to distinguish source $\gamma = 1$3, although the overall Spearman correlation between slide quality and AI-guess rating was $\gamma = 1$4 (Leinonen et al., 13 May 2026).

6. Benchmarks, failure modes, and open questions

Benchmarking is central because many AI-PPP failures emerge only in multi-turn, multimodal settings. PPTC evaluates PowerPoint Task Completion over 279 multi-turn sessions—229 for creating new PPT files and 50 for editing existing long templates—using PPTX-Match, which compares the final PPT output rather than the label API sequence. It defines

$\gamma = 1$5

and

$\gamma = 1$6

with success requiring exact non-position matches and satisfaction of all position-related rules. GPT-4 reaches 75.1% turn-based accuracy but only 22.7% session-based accuracy for new PPT creation, and 38.1% turn-based versus 6.0% session-based accuracy for editing templates, with position-related accuracy at only 24% (Guo et al., 2023).

Other evaluation frameworks broaden the target beyond API correctness. SlidesBench, introduced for natural-language-to-slide generation, supports both reference-based metrics—element matching, content similarity, color similarity, and position similarity—and reference-free GPT-4o judgments of text, image, layout, and color quality on a 0–5 scale; the paper reports human–model agreement of ICC = 73.8%–85.3% and finds that iterative refinement improves quality in all three instruction settings (Ge et al., 1 Jan 2025). PPTEval rates Content, Design, and Coherence on a 1-to-5 scale and reports average Pearson 0.71 and Spearman 0.74 correlation with human judgments, while DeepSlide’s dual-scoreboard explicitly separates Artifact Scoreboard from Delivery Scoreboard so that narrative flow, pacing precision, slide–script complementarity, attention choreography, and rehearsal readiness are not collapsed into static deck quality (Zheng et al., 7 Jan 2025, Yang et al., 1 Apr 2026).

Recurring limitations are consistent across the literature. Live assistance is constrained by latency, ambiguity, and difficulty inferring complex transformations; notebook systems report retrieval errors, long-context limits, and inability to access some kernel-state findings; edit-based generators still struggle with intricate nested group shapes, overlapping elements, and imperfect success rates; delivery-oriented systems note missing human evaluation for audio modules or the absence of dynamic animations; and coaching systems expose transcription error, selective uptake of recommendations, and unresolved links between formative feedback and final performance (Takahira et al., 16 Oct 2025, Wang et al., 2024, Zheng et al., 7 Jan 2025, Aggarwal et al., 11 Jan 2025, Shi et al., 5 Jul 2025, Becerra et al., 6 May 2026, Higham et al., 3 Oct 2025). A persistent controversy concerns autonomy: NB2Slides participants explicitly questioned full automation, and later systems repeatedly preserve presenter, instructor, or student control through outline editing, accept/reject decisions, interactive revision, or selective adoption of advice (Zheng et al., 2022). This suggests that the dominant research trajectory is not autonomous slide authorship in isolation, but mixed-initiative presentation engineering grounded in retrieval, executable representations, audience adaptation, and delivery-aware evaluation.