Visual presentations

A visual presentation turns text or data into a short sequence of slides or panels, so that a reader understands and remembers it better than from running prose. The design questions are how to split the content into a sequence, how to make each panel’s main point clear, whether the visual quality of the result matters, how to turn prose about relations or processes into a diagram, and whether software can do the conversion. The main disputes are whether polish and preference predict learning (they largely do not), whether visual formats beat text at all for retention, and how reliably generated slides carry the source’s content.

History

Sources use genre names for short sequences of visuals, not one fixed vocabulary. Segel and Heer reviewed news and research examples and identified distinct genres of narrative visualization (Segel & Heer, 2010). A later systematic review describes slideshows as a way to explain data to an audience, not to analyze it; it defines infographics as combining graphics, images and text, and scrollytelling as combining storytelling and scrolling, and calls data comics an upcoming genre (Schröder et al., 2023). Zhao and Elmqvist argue that data storytelling underuses the available media and propose a taxonomy of media types (Zhao & Elmqvist, 2022).

The history of slide design and of the multimedia-learning principles behind it is not covered by the research on file.

Core concepts

Segmenting content into a sequence

Presenting a multimedia message in learner-paced segments, not as one continuous unit, improves deeper learning; Mayer and Pilegard report support in 10 of 10 experimental tests, with a median effect size of 0.79 (Mayer & Pilegard, 2014). The explanation offered is essential overload: lessons that are fast and complicated exceed what the learner can process at once (Mayer & Pilegard, 2014).

A common limit of about four items per panel borrows Cowan’s estimate that working-memory capacity is about four chunks, an inference not tested on slides (Cowan, 2001); see Working memory.

For the order of panels, Hullman and colleagues analysed 42 professional narrative visualizations and propose ranking transitions between visuals by their cost to the audience, to choose a linear, slideshow-style order (Hullman et al., 2013). Sequences that use parallelism, repeated structure from one visual to the next, gave memory and subjective-rating benefits (Hullman et al., 2013). Narrative sequencing as a technique is the subject of Visual storytelling.

Making each panel’s point clear

The best-tested device is the headline that states the takeaway as a sentence. In a 110-student experiment, slides built on multimedia principles with sentence headlines gave better comprehension and fewer misconceptions than default PowerPoint slides, and lower perceived cognitive load (Garner & Alley, 2013). A summary of the same study reports that the sentence-headline group scored higher on every question type of a delayed quiz (Nordell, 2016). In four large lecture sections, students who saw sentence-headline slides averaged 79% correct against 69% for traditional phrase-headline slides (Alley et al., 2006).

The advantage was uneven: it was significant on seven of fifteen questions, and sentence-headline students scored significantly lower on two (Alley et al., 2006). The 110-student contrast also bundled several design features, so the headline alone is not isolated. Farkas argues that the control slides were poor examples of standard design, so the experiments overstate the gain, and that the case for sentence titles weakens when a speaker can state the main idea aloud (Farkas, 2010). That second point does not apply to a panel read without a speaker.

Titles also steer what readers remember. In an eye-tracking and recall study, titles and supporting text should convey the message of a visualization, and redundancy between data, text and message helped communicate it (Borkin et al., 2015). Press coverage of that study says that strong visualizations have titles summing up the takeaway, that many titles add nothing, and that readers spend the most time on text, above all the title (Conner-Simons, 2015). The steering has a cost: viewers’ recall of a message aligned more often with a slanted title than with the visualization itself, and most still perceived the visualization as impartial (Kong et al., 2019). A headline must therefore match its visual.

Annotation works within limits. When chart and caption both emphasise a feature, readers take it as the takeaway; a caption pointing at a low-prominence feature loses to the visually prominent one (Kim et al., 2021). Text describing statistical or relational components of a chart steers takeaways toward statistics or comparisons, and the best placement (title or near the data) depends on the kind of information (Stokes et al., 2022). Heavily annotated line charts were not penalized and readers preferred them (Stokes et al., 2022). Captions help readers memorise the features they describe, but in one paper a caption did not significantly increase takeaways about the primary feature of single-line charts (Zhu et al., 2022). Signaling, cues that link text and picture, gave a small-to-medium gain in transfer and comprehension in a meta-analysis of 27 studies, mainly for learners with low prior knowledge (Richter et al., 2016). Perception of single charts is covered in Information visualization.

Text on a slide read alone

Social scientists averaged 55 words per slide, and the researcher notes that audiences process pictures with speech more easily than text, a spoken context (Wageningen University & Research Centre, 2015). Practitioners distinguish a self-read slide document, with words and images paired, from a projected deck, but give no tested limit (Duarte, n.d.).

Diagrams and charts from prose

Few classifies quantitative messages into seven relationship types (nominal comparison, time series, ranking, part-to-whole, deviation, frequency distribution, correlation) and gives a graph form for each, assigning bars to comparison and ranking, lines to time series and a scatter plot to correlation; this concerns numbers, not prose, and reports no testing (Few, 2004). When participants sketched graphs to fit verbal descriptions, they tended to use bars for discrete comparisons and lines for trends (Zacks & Tversky, 1997).

Systems that convert prose follow an extract-then-render pattern. ChartifyText first infers a table from data-involved prose with a language model, then generates a chart with added encodings and concise text (Zhang et al., 2024a). For processes, large language models can extract process-model information from text, and the number of example texts, the specificity of definitions and the rigour of format instructions are key to accuracy (Neuberger et al., 2024). A three-stage pipeline extracts a graph, checks its structure, then checks that the flow logic matches the wording of the text (Ying et al., 2026). Doc2Chart picks a chart type with a heuristic-guided module and beat baselines on chart-type accuracy, scored by automated metrics, with no reader test (Jain et al., 2025). A benchmark of model-generated Mermaid sequence diagrams shows significant capability gaps across models (Shbita et al., 2025).

Questions, controversies, and new developments

Does visual design quality matter?

Polish and embellishment appear to act on credibility and memory more than on measured comprehension. Some chart embellishments raise perceived credibility and others lower it; hand-drawn fonts and cartoon-style bars lowered it (Song et al., 2025). Aesthetically appealing infographics and visualizations in social media posts were rated more credible, with processing fluency helping to explain the effect, yet production quality did not change credibility ratings and data visualizations were not rated more credible than text (Computational Multi-Modal Communication Lab, n.d.). A PsyArXiv preprint reports that beauty causally raised trust in graphs, while actual misleadingness did not (Lin & Thornton, 2021). Trust in a visualization has several dimensions, and ratings depended on design and source (Pandey et al., 2023).

For memory, embellished charts were described as accurately as plain ones and recalled better after two to three weeks (Bateman et al., 2010), and colour and recognisable human-related objects raised memorability (Borkin et al., 2013). Font disfluency did not change paragraph reading comprehension in one poster (Tahiyah et al., 2024). Most of this evidence is on single charts, not slides, so claims about multi-panel polish go beyond it.

Preference versus learning

In these studies, what readers liked did not predict what they learned. In a lecture experiment with 237 students, text slides gave significantly higher immediate test scores than visualizations (64.1% against 59.5%), although the visualization condition was rated better in slide quality; a week later the difference had disappeared, while slides still beat no slides (Blokzijl & Andeweg, 2006). An English report of the experiment records that listeners significantly preferred visual slides although they learned at least as much from text slides (Blokzijl & Andeweg, 2007). Learners rated verbatim-redundant slides as giving the best understanding and engagement, although non-redundant slides gave better applied comprehension (Fenesi & Kim, 2014). Learner modality preference gave no main effect on comprehension, only an interaction (written-text preference with written text) (Lehmann & Seufert, 2020). A preprint finds a similar split for immersive training, which raised learners’ own assessments of learning without improving outcomes against a static diagram (Clegg et al., 2024). The pooled evidence is a few small experiments in different formats, so the pattern is suggestive, not settled.

Panels, scrolling and reveal

For click-through panels against scrolling, one crowdsourced experiment (180 participants) found a slideshow layout of text plus visualizations gave significantly better comprehension than a vertical layout, with better recall only when text and chart were linked (Zhi et al., 2019). Within a slide, results split. In a lab study of narrated slides, revealing visuals cumulatively in step with the voiceover gave higher post-test scores than showing the whole slide at once (Ito & Ichikawa, 2026). In Mahar and colleagues’ study (N = 93), students who saw static slides scored higher than those who saw slides that revealed content by animation (81.11% against 65.81% in one table; the body text reports 81.98% against 71.43%) (Mahar et al., n.d.). A meta-analysis found animation beat static graphics by a small overall effect, larger when system-paced (Berney & Betrancourt, 2016). Scrollytelling matched other formats on comprehension while improving experience, with higher engagement and lower cognitive load than text in one preprint (Mendez & Such, 2026), and static, scrollytelling and chatbot onboarding did not differ in comprehension in another (Chen et al., 2026a).

Visual summaries against text

LLM-generated tables and mind maps cut comprehension time by 42.9% and 31.9% against text with no loss of accuracy, though the authors note models struggle to generate structured output and need prompting strategies and critics for factuality and structure (Jain et al., 2024). Infographic summaries of medical articles were preferred (mean 3.97 on a 1-9 scale where 5 is neutral) and needed less mental effort (4.30 against 5.06), but delayed retention did not differ (Martin et al., 2019). A graphical overview read at the start of a hypertext improved comprehension of difficult hypertexts, while one read at the end lowered comprehension of easier ones; the field has no consensus (Salmeron et al., 2009). Students who read a science comic about fish neuroscience gained as much understanding as journal-article readers and showed the largest positive attitude shift (Science Societies, 2024).

Generated slides

Slide generation was defined as a task combining summarization, image and text retrieval, and layout prediction, with new metrics and human evaluation, in DOC2PPT (Fu et al., 2022). PPTAgent generates presentations successfully in most runs with strong models, but its authors report design flaws such as overlapping elements and say it does not use visual feedback to refine slides; its judge agrees with humans least on coherence (Pearson 0.55) and most on design (0.90) (PPTAgent authors, 2025). Newer benchmarks call holistic judge evaluations too coarse, and PresentBench reports NotebookLM beating other generators (Chen et al., 2026d).

The most documented error is content. In SlidesGen-Bench about one in five quiz questions answered from generated slides was wrong; of those, 66.3% came from missing content and 31.7% from mismatched values, and top systems omit less but mismatch more (Yang et al., 2026). Its aesthetic scores agree with human ratings better than a judge baseline (Spearman 0.71 against 0.57), still below human-to-human agreement (0.85) (Yang et al., 2026). Strong commercial generators often produce decks without structural features such as slide masters, and no system scored well on both visual quality and editability (Yang et al., 2026).

The benchmarks measure coverage and appearance, not learning. Students rated AI-generated slides about as good as instructor slides and could not reliably tell them apart, with higher quality ratings going with lower “AI-generated” ratings; the measures were ratings, not test scores (Leinonen et al., 2026). A study of 24 researchers compared Auto-Slides decks to chat-based reading on self-report understanding and memory, not a knowledge test (Yang et al., 2025).

Broader connections

Segmenting rests on the multimedia-learning view that overload limits comprehension, which links this topic to Working memory and to Writing for comprehension. Chart perception and the choice between prose, table and chart belong to Information visualization; the use of narrative order to introduce a dense document belongs to Visual storytelling; layout for screens belongs to Interface design.

Further reading

• Mayer and Pilegard on segmenting (Mayer & Pilegard, 2014): the handbook chapter on learner-paced segments.
• Garner and Alley on assertion-evidence slides (Garner & Alley, 2013): the sentence-headline experiment, best read with Farkas’s critique (Farkas, 2010).
• Bateman and colleagues on embellishment (Bateman et al., 2010): a test of whether decoration helps memory.
• Hullman and colleagues on sequence (Hullman et al., 2013): transitions and parallelism in narrative visualization.
• Yang and colleagues on SlidesGen-Bench (Yang et al., 2026): the error profile of generated slides.