NeurIPS 2026 — visual design gallery
Read the display, then its design lesson. Tables and numeric charts below reconstruct explicitly attributed factual subsets. Diagrams are original teaching layouts, not screenshots or reproductions of the papers' architecture. Selected examples are not a prevalence ranking of visual styles.
Controlled-ablation table with seed variability
PhaseLoRA: Control-Regime-Conditioned Low-Rank Adaptation for Continuous-Action Vision-Language-Action Policies · Table 3, p. 9; Section 4.4, p. 8; Appendix A.2, p. 14
| Control / method | Success rate (%) ↑ |
|---|---|
| Matched-parameter high-rank LoRA | 71.0 ± 1.0 |
| Random descriptors | 66.3 ± 3.2 |
| Scalar-gated update | 73.7 ± 1.5 |
| Without proxy supervision | 78.7 ± 3.1 |
| PhaseLoRA | 85.3 ± 1.2 |
Observed in the source: Choose controls that challenge the claimed explanation: added capacity, arbitrary temporal modulation, scalar rescaling and descriptor supervision are different alternatives.
Prepare: Choose controls that challenge the claimed explanation: added capacity, arbitrary temporal modulation, scalar rescaling and descriptor supervision are different alternatives.
Do not infer: Selected five-row subset, not the whole eight-row table. Standard deviation is not a confidence interval or a significance test. These are source-reported values, not an independent reproduction.
Nested-estimator calibration bars with a target reference line
Computer Use at the Edge of the Statistical Precipice · Figure 1b and Section 6.1, p. 7; Appendix M, pp. 21–24
Observed in the source: Place a nominal target on the chart and isolate which resampling level repairs undercoverage.
Prepare: Place a nominal target on the chart and isolate which resampling level repairs undercoverage.
Do not infer: Rounded labels transcribed from the rendered figure. The simulated setup is calibrated to DigiWorld evaluations; these are not universal coverage values or model task-success scores. The source figure shows no uncertainty bars for these three aggregate labels.
Three-stage reconstruction: initialize, then freeze
TransmissiveGS: Residual-Guided Disentangled Gaussian Splatting for Transmissive Scene Reconstruction and Rendering · Figure 3, p. 5
Observed in the source: Three vertically arranged stages match the text's training schedule. Solid and dashed arrows distinguish computation from gradients. Initialization and freezing are labeled rather than left implicit. Intermediate residual, normal and image outputs connect abstract modules to measurable physical quantities.
Prepare: Use one row per optimization stage; label initialization, frozen components and the image composition. In the manuscript, pair this map with actual normal, residual and rendered-image panels, as Figure 3 does.
Do not infer: An original explanatory redraw must retain stage order and the frozen geometry/scattering relationships; copying the topology without these qualifiers would change the method.
Assumption-labeled theorem/equivalence map
Unifying Goal-Conditioned RL and Unsupervised Skill Learning via Control-Maximization · Figure 3, p. 6
Observed in the source: Nodes are distinct mathematical formulations, not merely model names. Every bridge is conditional and points to a proposition. The legend distinguishes identical policy orderings from shared optimal policies.
Prepare: Use the named distinctions in the source when choosing labels, paths and explanatory panels for your own argument.
Do not infer: A simplified map must preserve the two different equivalence meanings and the assumptions attached to edges; do not relabel it as empirical algorithm performance.
Controlled-ablation table with between-subject variability
MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding · Table 2 and pretraining-ablation discussion, p. 7
| Masking ratio | Top-1 retrieval accuracy (%) ↑ |
|---|---|
| 0.15 | 52.58 ± 8.06 |
| 0.30 | 53.93 ± 6.76 |
| 0.50 | 52.40 ± 6.59 |
| 0.75 | 52.85 ± 7.51 |
Observed in the source: Hold the decoder fixed, expose the remaining hyperparameter and retain variability. The paper notes that the mean differences are within the standard deviations instead of presenting the best cell as decisive proof.
Prepare: Hold the decoder fixed, expose the remaining hyperparameter and retain variability. The paper notes that the mean differences are within the standard deviations instead of presenting the best cell as decisive proof.
Do not infer: One selected decoder row from a larger ablation. Between-subject standard deviation is not a confidence interval or a formal test of configuration differences.
Task-chain diagnostic comparison, with explicitly different outcome definitions
SALART-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated Images · Section 4.1.1 and Figure 4, p. 7; main result Table 1, p. 6
| Outcome criterion (same model) | Percentage |
|---|---|
| Q1 artifact-presence recall | 99.4% |
| All four artifact-side questions correct for an image | 53.3% |
Observed in the source: Show that a high marginal detection score can coexist with much lower success on an entire grounded question chain. Label the marginal and joint definitions directly on the visual.
Prepare: Show that a high marginal detection score can coexist with much lower success on an entire grounded question chain. Label the marginal and joint definitions directly on the visual.
Do not infer: Do not label both bars generic accuracy or treat their difference as a model-to-model gain. The source's four-question success is joint, not the mean of question accuracies. No uncertainty interval for this small subset was established in the scan.
Caption and visual checklist
Section titled “Caption and visual checklist”State the display’s question or supported finding, define panels and symbols, keep conditions and units visible, and explain uncertainty. Label schematic curves as schematic. Separate training-only paths from inference. Retain baseline wins and preserve metric direction. View the final display at ordinary PDF size and in grayscale; do not rely on color alone.
Use the shared visual blueprints for additional architecture, proof, ablation and appendix-planning examples.