Visualization and figure design in scientific manuscripts

Last updated: 2026-05-17

Synthesis

Figure design has its own foundational literature, parallel to the sentence-level writing literature scriptorium already engages with. Tufte’s The Visual Display of Quantitative Information (1983; 2nd ed. 2001) [1] is the canonical statement of the design principles — maximise data-ink, minimise chartjunk, small multiples, data density, narrative integration with text. Wilkinson’s The Grammar of Graphics (1999; 2nd ed. 2005) [2] is the formal-system counterpart: a generative grammar for statistical graphics that underlies ggplot2 and virtually every modern visualisation framework. PLOS Computational Biology’s “Ten Simple Rules for Better Figures” (Rougier, Droettboom, Bourne 2014) [3] and Rolandi, Cheng & Pérez-Kriz’s Advanced Materials guide (2011) [4] are the practitioner-oriented translations of those principles into scientific-figure practice.

The under-researched problem in this space is figure-text alignment: does the figure caption actually describe what the figure shows, and does the body-text discussion of the figure match what the figure displays? Anecdotal evidence — visible in any thorough peer review — is strong that misalignment is common (wrong axis units called out, wrong panel referenced, claims about patterns the figure does not actually exhibit). Formal corpus studies of figure-text alignment in scientific articles are rare; the closest adjacent literature is image-forensics work on duplication and manipulation (Bik et al. 2016; see forensic-methodology), which addresses figure integrity rather than figure-text consistency.

For scriptorium specifically, this opens a candidate skill — figure-text-alignment — that is technically interesting but limited by current LLM capability to actually read scientific figures. Multimodal models can parse simple charts and extract axes and labels; they struggle with complex multi-panel figures, specialised plots (volcano plots, dimensional-reduction projections, Manhattan plots, gel images), and panel-level sub-references. The skill is worth scoping as v0.3 with explicit LLM-capability honesty: catch the cheap misalignments (panel-numbering, axis-unit mismatches, figure-counter drift) and route the harder ones to human review.

Evidence and frameworks

Tufte and the data-ink tradition

Tufte’s The Visual Display of Quantitative Information [1] introduced and codified the design principles that now define serious quantitative graphics:

• Data-ink ratio. Maximise the proportion of ink/pixels devoted to data; minimise non-data ink (heavy gridlines, redundant scale labels, decorative shading).
• Chartjunk. Decoration that adds no information actively degrades the figure.
• Small multiples. Arrays of similar small figures support comparison across conditions without imposing scale-switching cost on the reader.
• Graphical integrity. The visual representation must be proportional to the data; broken axes, truncated zero, and area-vs-length encoding errors mislead readers.
• Narrative integration. The figure exists to support a claim in the text; figures and text together are the unit of argument, not separate artefacts.

Tufte’s later books — Envisioning Information (1990), Visual Explanations (1997), Beautiful Evidence (2006) — extend the framework but the 1983/2001 volume is the operational reference.

Wilkinson’s grammar of graphics

Wilkinson’s Grammar of Graphics [2] is a different kind of contribution: a formal generative grammar for statistical graphics. Any graphic is decomposed into:

• Data — variables and observations.
• Aesthetics — mappings from variables to visual channels (position, colour, size, shape).
• Geometries — the marks (points, lines, bars).
• Statistics — transformations of the data (binning, smoothing, summary statistics).
• Scales — mappings from data range to display range.
• Coordinate system — Cartesian, polar, etc.
• Facets — partitioning into small multiples.

The grammar is the intellectual substrate of ggplot2 (Wickham 2016), Vega-Lite, Altair, and virtually every modern declarative plotting framework. For scriptorium, the grammar is useful because it gives a structured vocabulary for describing what a figure should contain. A figure-text alignment skill that knows what the data, aesthetic mappings, and statistics of a figure are has specific things to check the text against; one that treats the figure as an opaque pixel array does not.

Practitioner translations

Rougier, Droettboom & Bourne (2014). [3] The most cited practical guide for scientific figures, in PLOS Computational Biology’s Ten Simple Rules series. The ten rules: know your message; know your audience; identify the right type of plot; captions are not optional; do not trust the defaults; use colour effectively; do not mislead the reader; avoid chartjunk; message trumps beauty; get the right tool. All ten map cleanly onto checkable properties of a figure-plus-caption pair, which is part of the reason the paper has the citation traction it does.

Rolandi, Cheng & Pérez-Kriz (2011). [4] An Advanced Materials brief guide co-authored by a scientist, a graphic designer, and a cognitive psychologist. The contribution is the cognitive-design overlay: figures should support the reader’s cognitive task, not just display the data. Cognitive-load framing (see hayes-flower-writing-model) applies in the visual modality — unnecessary distinctions in colour, irrelevant panels, distant legend-to-data placement, all add extraneous load.

Stephen Few, Information Dashboard Design (2nd ed. 2013). [5] The dashboard-design literature is adjacent but increasingly relevant as scientific figures move into interactive and dashboard-style forms (Shiny apps, IIIF viewers, paper-companion explorables). Few’s emphasis on pre-attentive processing and at-a-glance comprehension is operationally translatable.

Figure-text alignment

This is the under-studied area. The peer-review literature (critique-quality-evidence) consistently identifies figure-text misalignment as a high-frequency reviewer concern:

• Caption references the wrong panel (“Figure 2A shows X” when the X is in Figure 2B).
• Axis units in the body text disagree with axis units in the figure.
• Sample sizes claimed in text differ from those visible in the figure (numbers in bars, n= labels in caption).
• Discussion of a “trend” or “pattern” that is not visible in the cited figure.
• Figure counter drift — Figure 4 in the file is referenced as Figure 5 in the text after a mid-revision insertion.
• Colour-blindness violations (red/green encoding) when the caption asserts accessibility compliance.

Quantitative corpus work on figure-text misalignment specifically is sparse. The closest adjacent corpus work is in image-forensics (forensic-methodology), which catalogues duplication and manipulation but not text-misalignment, and in caption-quality studies (largely informal commentary). No formal figure-text- alignment corpus or prevalence study was located during this sweep; the absence is itself worth noting — practitioners assert these errors are common but, to our knowledge, no peer-reviewed prevalence estimate exists.

Image forensics overlap

Bik et al. (2016) [6] — already detailed in forensic-methodology — established that ~3.8% of biomedical papers contain inappropriately duplicated images. Most of those issues are integrity failures (panel re-use, contrast manipulation) rather than alignment failures (caption-figure mismatch). The two failure modes are conceptually separable:

• Integrity: the figure does not honestly represent the underlying data.
• Alignment: the figure and the surrounding text disagree.

Scriptorium’s natural focus is alignment, not integrity, because alignment is text-tractable and integrity requires image processing.

How this informs scriptorium

A figure-text-alignment skill is a v0.3 candidate, with explicit limits on what it can deliver.

What is tractable for an LLM-only skill:

• Cross-reference consistency. Does every Figure N reference in the body text correspond to a figure with caption “Figure N”? Does every figure-with-caption have at least one body-text reference? This is deterministic string-matching, not LLM judgement.
• Panel-letter consistency. Does “Figure 3B” in the text match a panel labelled “B” in Figure 3’s caption?
• Axis-unit and sample-size cross-check. If the caption says “n=24” and the body text says “twenty patients”, flag the discrepancy. Same for units (μg vs. mg, log-fold-change vs. fold-change).
• Counter drift. If figures are reordered, ensure body-text references are updated.

What is partially tractable with multimodal LLMs:

• Axis label match. If the LLM can read the figure, does the axis label match the variable the caption and body text describe?
• Plot-type match. If the caption says “boxplot” and the figure is a barplot, flag it.
• Colour-channel description. If the caption asserts colour coding by category and the figure uses gradient colour, flag it.

Multimodal capability is improving but is unreliable for: complex multi-panel figures, specialised scientific plot types (volcano plots, MA plots, Manhattan plots, k-mer spectra, gel/Western images), figures with embedded annotations and arrows, and any figure where extraction of panel-level data points is required. The skill must default to flagging rather than asserting, and must produce inspectable output where the LLM’s reading of the figure is exposed alongside its claim.

What is not tractable:

• Image integrity (duplication, manipulation). Belongs to Proofig / ImageTwin / human inspection; scriptorium should not pretend to do this.
• Pattern-claim verification (is the trend the author describes actually visible?). Requires quantitative inspection of the underlying data, not just the rendered figure.
• Statistical-claim verification from figures (is the p-value in the figure consistent with the rendered effect size?). Requires the underlying data.

Preservation contract. A figure-text-alignment skill that modifies text to fix references must respect the preservation constraints in semantic-preservation — particularly statistic-string preservation. A skill that “fixes” “Figure 4” to “Figure 5” must not also silently change “n=24” to “n=20” because the LLM thinks they should match.

Reporting. The skill’s output should be a structured list of candidate inconsistencies, each with:

• Location in body text (paragraph, sentence).
• Location in figure asset (panel, caption line).
• Type of inconsistency (counter drift, axis-unit mismatch, etc.).
• Confidence (high for deterministic checks, low for LLM-vision claims).
• Suggested fix (without auto-applying it).

Implementation priority for scriptorium

Verdict: Maybe later — figure-text-alignment is plausibly v0.3 but should not ship before its capability limits are honest.

Condition that would flip to Yes:

• Multimodal LLM capability on scientific figures becomes demonstrably reliable for the bounded sub-tasks listed above (axis-label match, plot-type match, panel-letter check).
• Author demand surfaces — the skill is most valuable to high-volume authors and to groups with many trainees producing draft manuscripts.
• The deterministic sub-skills (cross-reference consistency, panel-letter consistency, counter drift, axis-unit cross-check) are independently valuable and could ship earlier as a text-only first cut.

If Yes (anticipated scope):

• Skill name: figure-text-alignment
• Phase: v0.3 (after the conservative-edit posture and preservation reporting are well-established).
• Scope:
  • Sub-skill A (text-only, deterministic): cross-reference consistency, panel-letter consistency, counter drift, axis-unit/sample-size cross-check. Shippable earlier, possibly in v0.2.
  • Sub-skill B (multimodal): axis-label match, plot-type match, colour-channel description, gated behind a multimodal_enabled flag in MANUSCRIPT_STATE. Outputs are surfaced as low-confidence flags requiring human review.
• Required data:
  • Figure assets (PDF / PNG / SVG referenced from the manuscript).
  • Captions in machine-readable form.
  • MANUSCRIPT_STATE.figure_index listing every figure with its file path and caption.
  • For Sub-skill B, multimodal-capable model access.

Honest limits:

• Sub-skill B will produce false positives on specialised plot types (volcano plots, etc.). The skill must reserve flagging language for unambiguous cases.
• Image integrity is out of scope and must be named as such; the skill should explicitly cite Proofig / ImageTwin and direct authors to those tools (see forensic-methodology).
• The skill must not silently rewrite captions; misalignments surface as candidates for human resolution.

Open questions / weak evidence

• Formal corpus data on figure-text misalignment prevalence is sparse. Anecdotal evidence is strong but a peer-reviewed estimate of “what fraction of submissions have misaligned figure references” appears to be missing. What would close this gap: a corpus study analogous to Bik et al. (2016) for image duplication, but targeting caption-vs-body and axis-label-vs-body alignment.
• Multimodal LLM accuracy on scientific figures is improving rapidly but is not benchmarked specifically for the text-alignment task scriptorium would deploy. Internal evaluation is required before any Sub-skill B claims.
• The boundary between figure-text alignment and pattern-claim-verification is fuzzy. A claim that “Group A is higher than Group B” requires either trust in the rendered figure or independent access to the data; the latter is out of scope.

References

1. Tufte ER. The Visual Display of Quantitative Information. 2nd ed. Graphics Press; 2001. ISBN 9781930824133.
2. Wilkinson L. The Grammar of Graphics. 2nd ed. Springer; 2005. ISBN 9780387245447. doi:10.1007/0-387-28695-0.
3. Rougier NP, Droettboom M, Bourne PE. Ten simple rules for better figures. PLoS Computational Biology. 2014; 10(9):e1003833. doi:10.1371/journal.pcbi.1003833.
4. Rolandi M, Cheng K, Pérez-Kriz S. A brief guide to designing effective figures for the scientific paper. Advanced Materials. 2011;23(38):4343–4346. doi:10.1002/adma.201102518. PMID: 21960472.
5. Few S. Information Dashboard Design: Displaying Data for At-a-Glance Monitoring. 2nd ed. Analytics Press; 2013. ISBN 9781938377006.
6. Bik EM, Casadevall A, Fang FC. The prevalence of inappropriate image duplication in biomedical research publications. mBio. 2016;7(3):e00809-16. doi:10.1128/mBio.00809-16. PMID: 27273827.