Training · target available
- Receive input
xand targety. - Initialize a candidate prediction.
- Refine it using the energy gradient.
- Evaluate a supervised objective.
- Update model parameters through refinement.
ICLR 2026 — visual forms you can prepare
See the form before opening the source paper. These 12 examples show how to explain a mechanism, compare evidence or reveal a limitation. Each includes an inline illustration or attributed results table, an inspected oral-paper example, a preparation checklist and caption requirements.
The numerical OpenThoughts and FlashVID tables retain real source values and their comparison limits. Other drawings are original teaching illustrations, not screenshots of the original figures or experimental results. Hypothetical plots and schematic images are labeled prominently. They teach a visual grammar without redistributing unverified publisher artwork.
This is a selected gallery, not a ranking of the most frequent plot types. Figures should answer a question; an accepted paper’s style is not automatically one to imitate. Use the contribution-specific packages to choose what belongs in your paper and the 223-paper atlas to understand observed variation.
Architecture and representation flow
Best for: new architecture · self-supervised method · knowledge distillation · reinforcement learning.
Original teaching illustration · not the paper’s artwork or experimental results
What the inspected paper does: A left-hand encoder/predictor/target diagram links the current state, action and policy latent to a temporal-difference objective. The target branch visibly marks stop-gradient. A middle maze trajectory and right-hand representation cloud explain why the predictor summarizes policy-dependent future states rather than simply the next frame.
TD-JEPA: Latent-predictive Representations for Zero-Shot Reinforcement Learning · Figure 1, physical PDF page 2 · Full writing study
- Prepare
- One overview that reveals the changed computation; identify inputs, outputs, reused modules, learned modules, targets and gradient boundaries. Add an inset only where a novel operation would otherwise remain a black box.
- Caption checklist
- Define each symbol and arrow; identify whether the drawing shows training or deployment; state what quantity the prediction represents. Keep training-only targets visibly distinct from inputs available at inference.
- Avoid
- A teacher/student box diagram without losses, stop-gradient or input availability does not explain a distillation or SSL method. Do not interpret the cartoon's point clouds as measured embeddings.
Training versus inference pseudocode
Best for: optimization algorithm · iterative inference · new training objective · system method.
Original teaching illustration · not the paper’s artwork or experimental results
Inference · no target
- Receive
x; freeze model parameters. - Initialize multiple candidates.
- Refine each with a stated iteration budget.
- Compare their final energies.
- Return the selected prediction.
What the inspected paper does: Two compact algorithms are placed side by side. Both iteratively refine candidate predictions using the energy gradient. The training algorithm receives the target and updates model parameters through the refinement process; inference receives no target, repeats candidates, and selects the lowest-energy final candidate with model parameters fixed.
Energy-Based Transformers are Scalable Learners and Thinkers · Algorithms 1 and 2, physical PDF page 5 · Full writing study
- Prepare
- A short algorithm with named inputs, outputs, state, iteration budget and update rules; separate training and inference when their available information or mutable variables differ.
- Caption checklist
- Say which variables are optimized, which are frozen, what terminates the loop and whether gradients pass through inner updates. Define candidate count and iteration count separately.
- Avoid
- Listing mathematical updates without identifying the differentiation path can conceal the main implementation cost. Do not promise implementation equivalence without reconciling paper and code.
Geometric explanation beside a formal proposition
Best for: theory · identifiability · convergence analysis · method with a genuine guarantee.
Original teaching illustration · not the paper’s artwork or experimental results
What the inspected paper does: The diagram puts a matrix space above a curved distribution space and connects them by fibers and a projection. Corresponding highlighted paths show how a horizontal lifted path maps to a Wasserstein geodesic. The figure is beside a cited background proposition, not itself a proof of the new algorithm.
On the Wasserstein Geodesic Principal Component Analysis of probability measures · Figure 1 and Proposition 1, physical PDF page 3 · Full writing study
- Prepare
- For a theoretical contribution, place the precise assumptions and statement near a small construction, counterexample or dependency diagram. Give a proof sketch in the main text and a lemma-to-theorem map for the appendix.
- Caption checklist
- Define the spaces, maps and allowed domain; distinguish an explanatory illustration from a numerical experiment. Identify inherited facts versus new results.
- Avoid
- A visually plausible surface does not establish a theorem, and a theorem about an idealized model does not establish the practical implementation's guarantee. Do not add a theorem merely to make an empirical paper appear theoretical.
Benchmark table with comparison metadata
Best for: new method · dataset or recipe · benchmark · empirical comparison.
Real source values · attributed table reconstruction
| Model | Training examples | Training | AIME25 accuracy (%) | LiveCodeBench 06/24-01/25 accuracy (%) | GPQA-D accuracy (%) |
|---|---|---|---|---|---|
| OpenThinker3-7B | 1.2M | SFT | 53.3 | 51.7 | 53.7 |
| DeepSeek-R1-Distill-Qwen-7B | 800K | SFT | 38 | 34.5 | 33.2 |
| Nemotron-Nano-8B | 3.9M | SFT/RL | 48 | 50.9 | 52.9 |
What the inspected paper does: The table puts base-model family, training-set size, optimization method and data availability above task scores. Benchmarks are grouped by domain, with held-out tasks separate. Its bold values denote a two-standard-error band around the best, not simply the single highest number.
OpenThoughts: Data Recipes for Reasoning Models · Table 1, physical PDF page 2; evaluation protocol in Appendix B · Full writing study
- Prepare
- One principal comparison table with the closest baselines, model/data/budget conditions, metric direction and an explicit highlighting rule; retain rows or tasks where the proposed method does not win.
- Caption checklist
- State the task, split, score scale, evaluation recipe, variability and whether baselines were rerun or imported. Define every abbreviation and bold/underline rule.
- Avoid
- Different data sizes and training recipes do not form a controlled loss-function ablation. A table subset must not inherit a best-in-table claim from the complete original.
Ablation and sensitivity table
Best for: new architecture · loss or optimizer · token compression · distillation method.
Real source values · attributed table reconstruction
| alpha | VideoMME | EgoSchema | LongVideoBench | MVBench | Relative accuracy (%) |
|---|---|---|---|---|---|
| 0.0 / TSTM only | 56.7 | 60.2 | 55.3 | 55.6 | 97.4 |
| 0.2 | 56.2 | 59.8 | 55.3 | 56.5 | 97.4 |
| 0.4 | 56.4 | 60 | 55.1 | 57.2 | 97.9 |
| 0.6 | 57 | 60.4 | 55.8 | 57 | 98.5 |
| 0.7 | 57.8 | 60 | 56.5 | 57.4 | 99.1 |
| 0.8 | 57.2 | 60.1 | 56.3 | 57.1 | 98.8 |
| 1.0 / ADTS only | 56.9 | 60.1 | 55.6 | 57.6 | 98.5 |
What the inspected paper does: Two adjacent tables ask different mechanism questions: selection components in Table 4 and allocation parameter alpha in Table 5. Each repeats the same dataset columns. The sensitivity sweep retains the two endpoints representing one component alone, rather than showing only values near the best setting.
FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token Merging · Tables 4 and 5, physical PDF page 10 · Full writing study
- Prepare
- An ablation tied to each causal explanation plus a sensitivity sweep for consequential hyperparameters. Show the unchanged protocol and endpoints or controls needed to distinguish competing explanations.
- Caption checklist
- Define what is held fixed, exactly what each row changes, the full meaning of alpha and the metric scale. State whether a score is absolute or normalized to a baseline.
- Avoid
- A relative accuracy of 99.1% is not an absolute task accuracy of 99.1%. A parameter sweep is not proof that every component is necessary under all budgets.
Learning curves with intervention events
Best for: continual learning · optimization · training stability · data curriculum.
Original teaching illustration · not the paper’s artwork or experimental results
What the inspected paper does: A test-accuracy versus epoch plot keeps repeated reset drops visible and compares one and ten reinitialization iterations with an unchanged baseline. Two neighboring plots use iteration count to track approximation error and deviation from isometry, linking the behavioral trace to the proposed mechanism.
FIRE: Frobenius-Isometry Reinitialization for Balancing the Stability-Plasticity Tradeoff · Figure 6, physical PDF page 23, Appendix B.2 · Full writing study
- Prepare
- A trajectory rather than final-score bars when the contribution concerns stability, forgetting, recovery or convergence. Mark reset, schedule-change or intervention times, and accompany behavior with the relevant mechanism diagnostic.
- Caption checklist
- Identify the event, x-axis units, held-fixed model/data setup, seed count, smoothing and uncertainty. Explain whether variation across runs is shown; do not imply unshown error bands exist.
- Avoid
- Smoothing away a reset dip can remove the very failure the method aims to fix. Compare wall-clock cost as well as epochs if methods perform different amounts of work per step.
Measured points, fitted curves and extrapolation
Best for: scaling study · data efficiency · compute allocation · empirical law.
Original teaching illustration · not the paper’s artwork or experimental results
What the inspected paper does: Figure 5 first fits loss across finite ensemble sizes at each parameter count, then fits the resulting asymptotes across parameter counts. Figure 6 repeats a two-stage reduction across seed-token counts. A point on the second panel can therefore be an estimate from a previous fit, not a directly trained infinite ensemble.
Pre-training under infinite compute · Figures 5 and 6, physical PDF page 6 · Full writing study
- Prepare
- Show observed runs, the fit model, the fitted range and extrapolated region distinctly. If one estimate feeds another fit, draw or annotate that dependency instead of treating all markers as raw observations.
- Caption checklist
- Define every axis, compute/data constraint, fit equation and marker type. State whether points are measured, selected minima or fitted asymptotes and include uncertainty or sensitivity where available.
- Avoid
- An infinite-compute limit is extrapolated from finite experiments. A steeper fitted curve does not automatically mean better measured wall-clock efficiency or superiority at modern model scale.
Quality-cost scatter and Pareto frontier
Best for: efficient inference · agent benchmark · distillation · compression or systems.
Original teaching illustration · not the paper’s artwork or experimental results
What the inspected paper does: Five panels separate overall quality from task-category quality. Each plots score against dollar cost on a logarithmic x-axis, distinguishes agents with marker shape/color, includes error bars, and joins non-dominated operating points with a dotted frontier. The large shared legend is a readability tradeoff.
AstaBench: Rigorous Benchmarking of AI Agents with a Scientific Research Suite · Figure 2, physical PDF page 8; complete values in Table 4, page 18 · Full writing study
- Prepare
- An accuracy-quality versus actual cost, latency, memory or energy plot whenever efficiency is central; compare multiple operating points with the same resource denominator.
- Caption checklist
- Define whether cost is per answer, token, case or dataset; retain hardware/API pricing date, precision, batching, tool access, retries and aggregation where relevant. Explain uncertainty and the direction of improvement.
- Avoid
- FLOPs, prefill time and end-to-end latency are different denominators. A frontier under historical API prices is not a timeless model ranking; agent tooling and model versions also differ.
Matched heatmaps for controlled perturbations
Best for: robustness · ablation interactions · cross-domain transfer · benchmark diagnosis.
Original teaching illustration · not the paper’s artwork or experimental results
Method A · illustrative values
| Edges ↓ / features → | 0% | 25% | 50% |
|---|---|---|---|
| 0% | 90 | 76 | 62 |
| 25% | 78 | 64 | 50 |
| 50% | 66 | 52 | 38 |
Method B · illustrative values
| Edges ↓ / features → | 0% | 25% | 50% |
|---|---|---|---|
| 0% | 90 | 84 | 78 |
| 25% | 78 | 72 | 66 |
| 50% | 66 | 60 | 54 |
What the inspected paper does: Six annotated heatmaps compare two interaction modes across three datasets. The horizontal axis controls feature deletion, the vertical axis edge dropping, and one shared color scale encodes mean accuracy. Numeric cells permit local comparisons beyond the color impression.
Actions Speak Louder than Prompts: A Large-Scale Study of LLMs for Graph Inference · Figure 2, physical PDF page 7 · Full writing study
- Prepare
- A common-scale matrix when testing two nuisance factors, class confusion, transfer pairs or parameter interactions. Keep axes and ordering identical across methods and mark missing or untested cells distinctly.
- Caption checklist
- Define each perturbation operationally, including whether deletion is random or truncation; state dataset, repetitions, score direction and the shared scale. Summarize the conditional result rather than claiming one method wins everywhere.
- Avoid
- Per-panel auto-scaled colors can make incomparable values look equivalent. Here removing text features means truncation, while edge removal is random; calling both random deletion changes the experiment.
Qualitative gallery and paired editing panels
Best for: image generation · restoration · segmentation · vision-language editing · scientific imaging.
Original teaching illustration · not the paper’s artwork or experimental results
What the inspected paper does: The appendix groups images into generation, paired editing and free-form manipulation bands. Editing columns specify the operation and align before/after views, so the intended change is visible without searching the prose. The examples are selected qualitative demonstrations, not a success-rate estimate.
NextStep-1: Toward Autoregressive Image Generation with Continuous Tokens at Scale · Figure A1, physical PDF page 17 · Full writing study
- Prepare
- Input/reference/output rows with identical crop, order and display scale across methods; use task labels and zoomed regions that reveal the claimed detail. Include a declared selection protocol and complementary quantitative evidence.
- Caption checklist
- State the prompt or operation, checkpoint, sampling/selection rule and whether the panel is representative or selected. Describe what the examples establish and what they cannot quantify.
- Avoid
- A visually appealing montage alone does not show superiority. Do not compare different prompts, crops or favorable seeds without disclosure, and do not attribute later-checkpoint examples to the original model.
Data pipeline, attrition and multiplicity
Best for: dataset · benchmark · data recipe · synthetic-data distillation.
Original teaching illustration · rounded source counts; schematic widths
What the inspected paper does: A Sankey-like pipeline shows source groups flowing through question filtering, deduplication, subsampling and multiple-answer generation. Science, code and math remain distinguishable; a final multiplicative stage makes the difference between unique questions and answer examples visible. The original caption is very short and several pale labels are difficult to read.
OpenThoughts: Data Recipes for Reasoning Models · Figure 3, physical PDF page 9 · Full writing study
- Prepare
- A data-lifecycle figure for dataset or recipe papers: sources, selection, excluded records, deduplication, splits and final released unit, with licenses and leakage checks nearby.
- Caption checklist
- Define the counting unit at every stage and why counts grow or shrink. If flow widths encode quantity, say so and ensure they are proportional; otherwise explicitly call the diagram schematic.
- Avoid
- Questions, responses, tokens, images, patients and slides are not interchangeable counts. A final expansion stage is not evidence of additional independent source examples.
Failure gallery organized by mechanism
Best for: generative models · robustness · scientific imaging · deployment evaluation.
Original teaching illustration · not the paper’s artwork or experimental results
What the inspected paper does: The failure gallery uses rows for different visible artifacts: noise at the bottom, bottom-region solid color, whole-image noise and grid artifacts. The next figure separately compares NextStep-1 with NextStep-1.1; the latter is a changed checkpoint and must not silently replace the original evaluation.
NextStep-1: Toward Autoregressive Image Generation with Continuous Tokens at Scale · Figure A4, physical PDF page 22; read Appendix B and separate Figure A5 carefully · Full writing study
- Prepare
- A small taxonomy with one clearly annotated example per important failure mode, an explanation or hypothesis, and a measured frequency only if systematically counted. Pair attractive success galleries with these boundary cases.
- Caption checklist
- State model/version, condition and how examples were selected; distinguish observed artifact from hypothesized cause. Explain whether fixes require new training, preprocessing or extra inference cost.
- Avoid
- Selected failure examples establish existence, not prevalence. A later version's improvement does not retroactively improve the original model's benchmark results.