AAAI 2026 — visual examples

16 of 1058 oral records have a writing-focused manuscript study. Official program snapshot: 2026-10-25. Selected rendered pages were inspected for each study; this is not a proof audit or reproduction of experiments.

Teaching selection: 16 of 16 studies inspected. Choose an example by contribution style. Completing this selection does not mean the full oral inventory has been read.

Selected deep collection: the complete official main-track oral inventory is present, but only the counted studies have been read for writing analysis and visually inspected. Other records remain uninspected, not automatically scheduled or unavailable. This is not an all-paper AAAI writing scan.

The maintained teaching selection spans contrasting contribution styles and research areas. It is neither random nor statistically representative. Selected-study completion is separate from full-inventory coverage, and its measurements cannot establish conference-wide writing frequencies or a standard figure quota.

Main manuscripts use official OJS proceedings PDFs. Separately linked posters, code, extended versions and supplements are not counted as inspected appendices unless a study explicitly records a second inspected source. Formatting lessons use the 2026 call and author kit, not a later edition.

Inspect the complete inventory and source status · Official oral program

Choose a visual for its explanatory job

Use a diagram to explain a mechanism or assumption, a plot to expose a relationship, a table for exact comparison, and aligned qualitative panels for visible behavior. Captions should define the setting, symbols, units and supported conclusion. Match the display to the evidence, not to decoration.

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.

Theory roadmap: keep conditions beside each result

On the Learning Dynamics of Two-layer Linear Networks with Label Noise SGD · Table 1 and Condition 1 p.4; Phase II and Lemmas 5–6 pp.5–6; SAM extension pp.6–7

Argument segment	Setting or update rule	Claim and evidence role	Boundary to preserve
Phase I	Label-noise SGD in the specified two-layer linear setting	Weight-norm decrease and departure from the lazy regime; Lemma 3 / Theorem 2	Initialization, data and scale conditions remain attached
Phase II	Analysis switches to gradient descent after entering a small-weight regime	Alignment and convergence; Lemmas 5–6	Lemma 6 adds an alignment assumption; not the same noisy-SGD update
SAM extension	Empirical comparison	Related dynamical observations in matched-style plots	An analogy, not an additional proved theorem

Original source-guided roadmap. An original editorial roadmap combining Table 1 with the surrounding conditions and extension discussion. Wording is paraphrased; this is not a reproduction of the source table or a proof reconstruction. Each row has a different logical scope.

Observed in the source: The source's compact phase table aids navigation, but the body supplies consequential update-rule changes and assumptions that a compressed summary can hide.

Prepare: Put setting, operation, formal result and scope together. Use local proof intuition and a source locator rather than an unlabeled chain of theorem names.

Do not infer: Do not connect these rows into one unconditional theorem for noisy-SGD training of deep classifiers, or call the SAM observations a proved guarantee.

Theory-to-practice comparison: preserve unfavorable regimes

Generalization Bounds for Semi-supervised Matrix Completion with Distributional Side Information · Table 1 and Real Data Experiments, physical p.7

Dataset / method	RMSE at 90% labels removed ↓	RMSE at 95% labels removed ↓
Douban / IGMC	0.8713	0.8462
Douban / DAMC	0.8323	0.8577
Yelp / IGMC	1.2028	1.0859
Yelp / DAMC	1.129	1.139

Source-reported numerical subset. Selected two methods, two datasets and two of eight label-removal fractions from Table 1. Lower RMSE is better. The complete table includes other baselines and ML-100K. Real-data DAMC replaces the theoretical SVD stage with an autoencoder; no uncertainty is shown in this table.

Observed in the source: DAMC improves these displayed 90% settings, but IGMC wins both displayed 95% settings. The practical test illustrates the learning paradigm, not an unchanged execution of the theoretical algorithm.

Prepare: Group comparable conditions and show where a ranking reverses. State the practical departure from the theorem setting beside the result, not only in an appendix.

Do not infer: Do not transfer the synthetic experiment's 30-run protocol to this real-data table or infer universal sample-efficiency or recommender superiority.

Empirical resource sweep: a ranking depends on the operating point

Why Do Open-Source LLMs Struggle with Data Analysis? A Systematic Empirical Study · Figure 4, physical p.5; Reasoning Length pp.5–6

Both axes are linear; the vertical axis starts at zero. Lines guide the eye between the reported settings, not unmeasured budgets. Coincident points can overlap; use the exact table below.

Maximum tokens per turn	Original reasoning accuracy (%)	Full-trace accuracy (%)	Summary-trace accuracy (%)
1024	18.83	16.74	20.08
2048	15.9	15.9	17.15
4096	17.15	9.62	18.83

Source-reported numerical subset. Replotted DiscoveryBench panel of Figure 4; exact token maxima come from the surrounding prose rather than rounded tick labels. The original/full traces share the middle value. Only this dataset panel is shown, and the source supplies no error bars or repeated-run uncertainty. Trace content also changes, so this is not a pure token-cost ablation.

Observed in the source: At the largest shown budget, full traces fall to 9.62% while summarized traces reach 18.83%. A single favorable point would hide this operating-condition dependence.

Prepare: Use a named resource axis, preserve all selected settings and state what else changes. Provide the exact numeric table and comparison protocol alongside the plot.

Do not infer: Do not infer a causal attention mechanism, universal reasoning hierarchy, significance or outcomes between the measured budgets from the connecting lines.

Architecture: show the information-removal boundary

URaG: Unified Retrieval and Generation in Multimodal LLMs for Efficient Long Document Understanding · Figure 3 and Methodology, physical pp.4–5

Original teaching construction. Original simplified drawing based on URaG Figure 3, not the authors' artwork. Neutral blocks denote the reused backbone; the accent branch scores contextualized page/query tokens and controls pruning. Crossed-out token boxes are discarded; every query token continues. The token counts are illustrative, not literal page counts: the implementation selects k=5 pages after layer 6 by default. The separate training inset does not supply ground-truth pages at inference. Projector, normalization and loss details are omitted.

Observed in the source: The source's token diagram makes the removal of unselected visual-page tokens visible, while Methodology specifies staged retriever training and joint LoRA training. Early layers still process the full input; savings occur downstream of selection.

Prepare: Show the exact pruning boundary, what survives, and the computation that remains. Distinguish data-flow arrows from selection control and keep training-only supervision outside the inference stream.

Do not infer: Do not imply all early visual computation is skipped, that ground-truth page labels are available at test time, or that operation-count savings establish measured latency.

Matched training controls: preserve a full-model non-win

URaG: Unified Retrieval and Generation in Multimodal LLMs for Efficient Long Document Understanding · Table 4, physical p.7

Variant	SlideVQA EM ↑	MMLongBench-Doc GACC ↑	LongDocURL GACC ↑
Qwen2.5-VL-3B	59.1	25.5	40
Baseline with SFT	61.9	29.1	37.3
URaG-3B without MLLM fine-tuning	62.1	29.4	43.1
URaG-3B	63.8	31.1	41.5

Source-reported numerical subset. Four-row Table 4 comparison at the source's reported score scale. EM is exact match; GACC is Generalized Accuracy. Higher is better. No uncertainty is shown. Without MLLM fine-tuning does not mean an untrained retriever.

Observed in the source: The controlled variants separate backbone fine-tuning from the retrieval intervention. The full system does not beat its no-MLLM-fine-tuning variant on LongDocURL (41.5 versus 43.1).

Prepare: Include the baseline with the same relevant fine-tuning opportunity and keep every comparator's input and update contract visible.

Do not infer: Do not combine EM and GACC into an unqualified average, infer significance, or highlight the full row as uniformly best.

Representation controls: connect module names to separate domains

VAEVQ: Enhancing Discrete Visual Tokenization Through Variational Modeling · Table 3, physical p.7; module definitions pp.3–5

Variant	VLQ	RCS	DCR	ImageNet rFID ↓	BraTS24 rFID ↓
Baseline	Off	Off	Off	8.02	10.47
M1	On	Off	Off	2.84	5.12
M2	On	On	Off	1.92	2.98
M3	On	Off	On	2.18	3.26
Full	On	On	On	1.14	2.5

Source-reported numerical subset. Exact switches and reconstruction FID from Table 3. VLQ: variational latent quantization; RCS: representation coherence strategy; DCR: distribution consistency regularization. Lower rFID is better; no uncertainty is shown. BraTS24 is an MRI-slice evaluation, not WSI.

Observed in the source: The architecture's named components become explicit switch columns, and the two datasets remain separate rather than disappearing into a single gain.

Prepare: Include partial combinations that distinguish local and global interventions, with metric direction and domain named beside the numbers.

Do not infer: Do not infer additive causal effects, significance, pathology applicability or clinical utility from these means; reconstruction FID is not generation FID.

Pathology controls: a metric-specific rather than universal win

Graph-Semantic Guided Learning for Virtual Immunohistochemistry Staining on Consecutive Histology Sections · Table 2 and Evaluation Metrics, physical pp.5–6

GSRM	DBD	FID ↓	KID ↓	SSIM ↑	DAB histogram KL ↓
Off	Off	57.608	2.101	0.243	3.935
On	Off	45.372	1.419	0.237	2.128
Off	On	42.326	1.109	0.251	3.438
On	On	38.292	0.801	0.248	1.966

Source-reported numerical subset. MIST component ablation from Table 2 at the source's reported scales. GSRM is graph-semantic rectification; DBD is the dual-branch discriminator. No repeated-run uncertainty is provided. The reference sections are weakly paired, not pixel-identical tissue.

Observed in the source: DBD alone has higher SSIM than the full method, while the full combination improves the other displayed measures. The main text qualifies SSIM because consecutive sections do not have exact spatial correspondence.

Prepare: Name what each image metric measures and explain the acquisition mismatch before interpreting its direction or rank. Keep the metric-specific non-win.

Do not infer: Do not make the whole final row 'best', treat image-distribution distances as biomarker accuracy, or present staining similarity as validated diagnostic substitution.

Scientific input contract: predicted is not absent

MSAnchor: De Novo Molecular Generation from Mass Spectrometry Data with Anchor-Extended Molecular Scaffolds · Table 2 and molecular-formula ablation, physical p.7; Table 1 p.5

MassSpecGym variant	Top-1 exact recovery (%) ↑	Top-1 Tanimoto similarity ↑
Without conditional information bottleneck	2.44	0.29
Without anchor-extended scaffold	2.28	0.27
Without side-chain prior	2.35	0.28
Predicted formula (source: w/o Formula)	1.72	0.26
MSAnchor with ground-truth formula	2.68	0.32

Source-reported numerical subset. Top-1 subset of Table 2. The source's 'w/o Formula' row still uses a MIST-CF-predicted formula; it is not formula-free. No uncertainty appears in this ablation table. The omitted full-model top-10 similarity differs between Tables 1 and 2 (0.41 versus 0.42), an unresolved source inconsistency.

Observed in the source: The formula comparison changes input quality, not merely a network component. Exact recovery and softer structural similarity provide different views of success; the absolute top-1 recovery remains low.

Prepare: State which information is known, predicted or absent. Show exact task success beside softer similarity and explain any conflicting reported values.

Do not infer: Do not rename this variant 'no formula', silently harmonize the top-10 conflict, or equate molecular similarity with experimentally confirmed identification or therapeutic benefit.

Robot evaluation: keep successes, trials and missing entries

FoAM: Foresight-Augmented Multi-Task Imitation Policy for Robotic Manipulation · Table 2, physical p.7; evaluation protocol pp.5–6

Policy	Scenario I	Scenario II	Scenario III	Scenario IV	Unseen task
MT-ACT	6/40	10/40	10/40	10/20	0/10
Gimg-ACT	7/40	8/40	11/40	Not reported	1/10
FoAM without foresight augmentation	7/40	13/40	13/40	12/20	1/10
FoAM	11/40	11/40	17/40	14/20	3/10

Source-reported numerical subset. Real-robot successes/trials from Table 2. Scenarios: I pick a specified tube; II insert a tube in a specified hole; III place a specified fruit/vegetable in a bowl; IV place bitter melon on a specified locker layer. The unseen case replaces the fruit with carambola. These are not the paper's simulation percentages or separately trained generated-goal experiments.

Observed in the source: The full model loses to the no-foresight variant in Scenario II, and one comparator has no Scenario IV result. Keeping fractions visible exposes the unequal and small trial denominators.

Prepare: Separate simulation, physical trials and qualitative demonstrations. Preserve missing entries, non-wins and the exact scope of each test.

Do not infer: A missing result is not zero, and these small counts are not precise deployment probabilities. The simulation headline's 86% versus 45% is a 41-percentage-point difference, not a 41% relative gain.

Component-control table: isolate the supervisory change

Modality-Balanced Collaborative Distillation for Multi-Modal Domain Generalization · Table 4, physical p.6

Components present	Average accuracy (%)
Baseline	56.47
EMA	60.94
Adaptive dropout + EMA	61.59
Gradient consistency + EMA	62.28
EMA + distillation loss	62.52
All four components	63.08

Source-reported numerical subset. Selected average column from Table 4, EPIC-Kitchens audio+flow multi-source domain generalization. Per-domain values and standard deviations are not reproduced here; the paper reports three seeds. The average column has no reported uncertainty interval.

Observed in the source: The table compares several individual additions on top of EMA, then their combination, instead of only contrasting a full model with the weakest baseline.

Prepare: Name the active components, keep input access and protocol matched, and report the outcome of each mechanistic control. Include uncertainty with an explicitly defined unit when available.

Do not infer: This six-row subset is not a full factorial experiment and cannot separate every interaction or prove component necessity in all settings.

Quality–latency plot: a cheaper proxy can still run slower

CATP: Contextually Adaptive Token Pruning for Efficient and Enhanced Multimodal In-Context Learning · Table 3, LLaVA-Next-7B block, physical p.7

Axes focus on the displayed range, not a zero baseline. Point numbers match the named configurations below the plot; exact values and comparison conditions remain in the table and caption.

Method	Relative performance (%)	Latency (s)	FLOPs (T)	KV cache (GB)
Vanilla	100	3.82	20.84	1.12
FastV	82.4	3.25	6.37	0.38
DivPrune	86.3	4.03	4.19	0.3
PLPHP	95.9	3.49	5.33	0.37
CATP	100.1	3.19	4.38	0.35

Source-reported numerical subset. One model block from the published Table 3: 4-shot inference, eight benchmarks, 77.8% pruning for pruning methods, NVIDIA H200 setup. Relative task performance is normalized to vanilla; it is not absolute accuracy. No uncertainty bars are supplied for this table. The table caption says four models although three blocks are displayed in the PDF.

Observed in the source: DivPrune has fewer reported FLOPs than CATP but greater latency than vanilla; plotting the measured endpoint makes that disagreement visible.

Prepare: Place quality and actual end-to-end latency on separate axes, retain the unpruned baseline, and report hardware, token budget, averaging and timing scope.

Do not infer: Do not infer universal speedups, raw accuracy or statistical significance from these point estimates; this is a factual subset, not a reproduced benchmark.

Architecture overview: separate deployed computation from teacher supervision

Modality-Balanced Collaborative Distillation for Multi-Modal Domain Generalization · Figure 2, physical p.3

Original planning layout. Original abstract drafting layout inspired by the explanatory role of MBCD Figure 2. It is not a reproduction of MBCD, omits its modality-specific branches, and should be adapted to the actual method.

Observed in the source: The source gives student/EMA teacher and modality-dropout internals separate visual regions, with a legend for features, heads and losses.

Prepare: Make model ownership and update versus supervision arrows explicit; give the changed operation a readable inset rather than enlarging every familiar encoder.

Do not infer: This teaching diagram is not an implementation specification or evidence of an observed result.

Put quality and elapsed cost in the same comparison

ProGMLP: A Progressive Framework for GNN-to-MLP Knowledge Distillation with Efficient Trade-offs · Table 3, physical PDF p.7 (printed p.24094); hardware/setup on physical p.5

Model	ogbn-arxiv accuracy (%)	ogbn-arxiv inference (ms)	ogbn-products accuracy (%)	ogbn-products inference (ms)
GCN	68.42 ± 0.67	84.13	72.94 ± 0.14	274.77
GCN + ProGMLP	70.11 ± 0.74	3.95	73.00 ± 0.12	52.21
GAT	69.25 ± 1.26	97.81	OOM	Not reported
GAT + ProGMLP	70.29 ± 1.94	12.8	Not reported	Not reported
GraphSAGE	69.70 ± 0.21	27.99	75.59 ± 0.10	139.21
GraphSAGE + ProGMLP	70.97 ± 0.85	6.93	75.68 ± 0.07	59.06

Source-reported numerical subset. Large-graph node classification in the published ProGMLP experiment. Accuracy is higher-better; inference milliseconds are lower-better. The main setup names an Intel i9-10980XE and NVIDIA A100 40GB. Displayed ± values are preserved as reported, without relabeling them as confidence intervals. OOM means out of memory; source dashes are written as Not reported.

Observed in the source: Paired rows expose the accuracy-latency trade-off and preserve failed or missing evaluations instead of hiding them behind a speedup headline.

Prepare: Place task, hardware, metric direction and missing-value meanings beside the table; report uncertainty definitions and timing boundaries explicitly in a new manuscript.

Do not infer: Do not confuse the number of executed students in a separate figure with milliseconds, infer performance for OOM cells, or repeat the source improvement row as percentage-point changes.

Separate correctness, coverage and conditional error

VCapsBench: A Large-scale Fine-grained Benchmark for Video Caption Quality Evaluation · Table 1 ALL column, physical PDF p.6 (printed p.12731); metric definitions and protocol on physical p.5

Caption model (source label)	AR (%, higher)	IR (%, lower)	CR (%, higher)
Qwen2VL-7B	47.5	9.7	52.5
Qwen2.5-VL-72B	63.1	11	70.9
GPT-4o	66.5	10.2	74.1
Gemini-2.5-Pro-Preview	76.7	10.3	85.5

Source-reported numerical subset. Selected four-model subset of VCapsBench Table 1's ALL summaries, using GPT-4.1 as the TextQA judge. The source defines AR and CR over all questions but IR over positive plus negative answers. These rounded reported aggregates are not raw question counts or independent rate measurements.

Observed in the source: The model with highest displayed accuracy and coverage does not have the lowest conditional inconsistency. Keeping all three columns prevents one number from standing in for every notion of caption quality.

Prepare: Write the metric numerator and denominator before the table, name the judge separately from the caption model, and retain trade-offs and annotation-model dependencies.

Do not infer: This selected subset is not the full ten-model, 21-dimension benchmark. Do not infer generated-video quality, confidence intervals or significance. The main paper acknowledges that Gemini-assisted QA construction may favor its coverage.

A completed run is not a successful solution

DeepProofLog: Efficient Proving in Deep Stochastic Logic Programs · Table 1, physical PDF p.6 (printed p.22386); interpretation in Section 6.1, physical pp.5-6

Variant	Digits N	Accuracy (%)	Training time (minutes)
DPrL (DP)	4	94.0 ± 0.3	25.4 ± 6.3
DPrL (PG)	4	93.5 ± 0.4	25.7 ± 8.6
DPrL (DP)	100	37.8 ± 2.9	105.8 ± 21.8
DPrL (PG)	100	0.0 ± 0.0	282.6 ± 20.5
DPrL (DP)	500	6.0 ± 4.9	469.0 ± 51.6
DPrL (PG)	500	0.0 ± 0.0	865.0 ± 67.9

Source-reported numerical subset. Selected DPrL rows and sequence lengths from the MNIST-addition experiment. DP is dynamic programming; PG is policy gradient. Accuracy concerns the full sum of two N-digit sequences. The source uses a 24-hour timeout for the wider comparison and reports the displayed uncertainty without its definition in this table caption.

Observed in the source: The table retains zero-accuracy completed runs. It therefore distinguishes computational completion from learning a useful policy, matching the adjacent acknowledgment of a suboptimal approximate policy.

Prepare: Pair accuracy and runtime, state the task unit and timeout, and give distinct symbols for timeout, out-of-memory, not measured and a genuine measured zero.

Do not infer: These are selected variants, not the full baseline table. Do not call a longer completed run successful reasoning or imply the ± quantities are confidence intervals.