Visual blueprints and caption workshop
These original examples demonstrate how to make an argument visible. They are neither copied paper figures nor measured results. For source-specific examples, use the ICLR 2026 writing-pattern studies.
Architecture: one intervention, two operating stages
Section titled “Architecture: one intervention, two operating stages”A useful architecture diagram exposes information access, the changed computation and the deployed path. It need not reproduce every layer of a familiar backbone. The first blueprint shows a generic distillation layout; the second shows how to distinguish measured points from extrapolation.
On small screens, the figures can be scrolled horizontally rather than shrinking labels beyond readability. Each also has a text description for assistive technology.
Caption workshop: architecture
Section titled “Caption workshop: architecture”Weak caption shape: “Overview of our proposed framework.” It identifies neither the computation nor the contribution.
A more informative original drafting shell:
Training and deployment of [method]. During training, [privileged input] provides [target] through a frozen [teacher]; [student] receives only [available input]. The proposed [operation] changes [specific relation]. Solid arrows denote [meaning], and dashed arrows denote [meaning]. At inference, only [modules] remain. Shapes/labels indicate [dimensions or state].
Fill this only with the actual method. If high-resolution processing remains in tissue selection, acquisition or another inference stage, it belongs in the deployment panel and cost accounting. “Low-resolution-only” is a claim about the full stated operating boundary, not a color choice in the figure.
Caption workshop: a result table
Section titled “Caption workshop: a result table”A precise caption identifies the comparison before interpreting it:
[Metric and direction] on [dataset/version/split], using [test-time input] and [evaluation unit]. [Rows] share [controlled resources]; [other rows] differ in [data/capacity/budget]. Values summarize [repetitions and statistic]. [Notation] marks [adapted or literature baselines]. The comparison supports [bounded observation]; [important exception] remains.
Do not invent an uncertainty interval to fill the template. If the source provides only point estimates, say that. For your own study, plan independent repetitions and uncertainty analysis before writing the caption.
A table layout that separates three scientific questions
Section titled “A table layout that separates three scientific questions”| Block | Question | Essential rows | Information that must not disappear |
|---|---|---|---|
| Main comparison | Does the method help in the stated setting? | Closest valid baselines and proposed method | Data access, split, metric, capacity and tuning budget |
| Mechanism ablation | Is the claimed component responsible? | Matched baseline, component variants and negative control | Which variable changes; whether retraining and budget remain comparable |
| Deployment comparison | What does the benefit cost? | Complete deployed pipelines at stated operating points | Hardware, precision, batch, preprocessing, auxiliary stages, memory and latency |
This layout need not become one huge table. Separate tables can be clearer, provided the text does not silently splice scores from incompatible protocols. See the FlashVID writing study and fixed-data distillation study for concrete protocol distinctions.
A diagram-to-algorithm consistency check
Section titled “A diagram-to-algorithm consistency check”For every arrow or module, locate the corresponding mathematical definition and implementation step. Check:
- The graph includes every input used by the algorithm.
- A stop-gradient, frozen parameter or moving average is identified explicitly.
- Losses supervise the intended module; a loss is not drawn as an inference output.
- An optional stage is labeled, and experiments specify whether it is enabled.
- A multi-step loop is not accidentally presented as a single operation.
- Pseudocode signs, normalization, sampling and stopping conditions match the implemented method.
The DTO-KD study illustrates why an exact derivation and a practical approximation need separate descriptions. The TD-JEPA study illustrates why illustrative pseudocode should be checked against the released implementation.
Appendix layout: a claim-indexed map
Section titled “Appendix layout: a claim-indexed map”An original appendix plan can use stable labels rather than a miscellaneous sequence of “additional” sections:
| Label | Purpose | Contents and forward links |
|---|---|---|
| A · Formal details | Verify the mathematical claim | Definitions, assumptions, proof dependencies, edge cases; link from each main result |
| B · Experiment contract | Reproduce the comparison | Data versions/splits, preprocessing, training, tuning, seeds, hardware and selection rules |
| C · Mechanism tests | Challenge the explanation | Controlled ablations, sensitivity, failed variants and competing explanations |
| D · Full evidence | Preserve coverage | Per-task/cohort results, uncertainty, examples and failure-selection policy |
| E · Artifacts | Find the implementation | Revision, environment, configurations, commands and table-to-output mapping |
The labels are a planning example, not required headings. A theory paper, dataset paper and clinical study need different contents. Keep essential method and validity information in the main text; the appendix supplies the audit trail, not a hidden replacement paper.
Publication-scale visual review
Section titled “Publication-scale visual review”Inspect the compiled PDF at its actual reading size, not only the source graphic at full-screen zoom. Check legend meaning, contrast, grayscale distinctions, font size, panel labels, crop context and consistent axes. Verify equations, table alignment and image resolution after the last compile. Finally ask a colleague to explain the main figure without your verbal assistance; revise the ambiguity they encounter.