AAAI — writing guide

Choose the track before planning the submission. AAAI covers a broad range of artificial intelligence research. The dates below belong to the main technical track; special tracks and workshops can have different schedules. Its two review phases are not two opportunities to submit the same paper.

The 2026 main-track oral inventory contains 1,058 papers across 213 sessions. Every record has a matched published manuscript. The teaching collection prioritizes deeply inspected examples across contribution styles and research areas. These are selected examples, not a representative sample or a completed all-paper scan. AAAI 2027 has planning dates, but is not being presented as an accepted-paper writing corpus.

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

Writing in brief

Start with the contribution, not a universal outline

AAAI covers methods, theory, empirical investigations, integrative work and critical analyses. State the problem → precise gap → intervention or argument → evidence → limits. A new objective needs its supervision and optimization explained; a benchmark needs a justified measurement target; a formal claim needs assumptions. Do not force every paper into an architecture-and-leaderboard story.

In MBCD’s study, modality imbalance motivates the corrective operations. In the recourse-fairness study, a measurement problem precedes the training proposal. These are examples, not frequencies across the oral corpus.

For theory, compare matrix completion’s assumption–algorithm–bound structure with the label-noise paper’s phase-to-result roadmap. State where model assumptions or update rules change. For an empirical investigation, the data-analysis-agent study organizes questions and controlled diagnostics before proposing a follow-up intervention; it also shows why every plot needs its own sample budget and interpretation limits.

Give each visual a question to answer

An overview should locate the changed computation; a zoom should explain its nonstandard operation. Label training supervision separately from inference inputs. Use matched plots for sensitivity, a table for exact comparisons, and pseudocode when iteration or update order would otherwise be ambiguous. A theorem is appropriate for a formal guarantee, not as decoration.

The inline gallery includes a teacher-supervision drafting layout, component controls and a quality–latency plot. CATP’s study shows why fewer FLOPs can coexist with greater latency. Caption labels count displays, not individual panels or experimental breadth.

For a changed architecture, compare FoundationSLAM’s feedback loop, URaG’s page-token selection and VAEVQ’s shared-decoder paths. Their explanatory jobs differ: show returned state, removed information or a changed representation rather than using the same generic block diagram for all three.

Make comparisons auditable and claims conditional

Keep model, dataset, split, input access, metric direction, budget and uncertainty meaning near the result. Separate external comparisons from component controls. Retain baseline wins. Use “we observe” for measured behavior and “we prove” only with the relevant assumptions and proof; local flatness, embeddings and correlations are not by themselves causal explanations.

Accepted papers can contain unclear captions or inconsistent notation. The studies retain such issues rather than silently turning them into exemplary conventions. Use the writing patterns and contribution-specific plans to choose what your own evidence package needs.

For applications, define the scientific outcome before choosing the metric. GSGStain teaches why consecutive tissue sections change image-comparison assumptions; MSAnchor distinguishes exact molecular recovery from similarity; FoAM separates simulated percentages from physical-robot successes and trial counts. None of those technical comparisons alone establishes unrestricted real-world utility.

Apply AAAI’s actual format

For AAAI 2026, the initial main-track allowance is seven technical pages, with references and the required reproducibility checklist outside that allowance. The kit uses two-column 10-point text and author-year citations. Ordinary figure and table captions both go below their displays; algorithm captions have a different convention. Optional supporting material cannot carry the only intelligible definition of the contribution.

Use the rules and official source register for precise phase-specific instructions, color and source-package restrictions, and unresolved source conflicts. Do not inherit CVPR/ICCV conventions or treat accepted PDF lengths as submission limits. No fixed figure count or drafting style guarantees acceptance.

Explore AAAI 2026

AAAI 2026

16 / 1058 inventoried oral papers have inspected writing studies. Snapshot: 2026-10-25.

Corpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

The searchable inventory includes both completed studies and clearly marked pending records. The atlas measures only inspected manuscripts; pending papers do not appear as fabricated zero-length or zero-figure records. For reusable drafting exercises, see argument and evidence and structure and visual design.

Choose a teaching example

Start with the contribution closest to your draft. Each linked study explains its argument, section order, visual choices, comparison conditions, supporting-material scope and wording. The type guides turn these observations into conditional drafting plans; the gallery shows inline examples. No paper count or display package guarantees acceptance.

16 of 16 selected studies inspected · 1042 oral records remain uninspected overall.

Purposive selection for contrasting explanatory jobs and research areas, not random sampling or a statistically representative survey. The initial seven studies are retained; nine additional manuscripts fill gaps in architecture, theory, empirical analysis and domain applications. Classification follows inspected arguments, not titles alone.

Architecture & representation (3)

• FoundationSLAM: Unleashing the Power of Depth Foundation Models for End-to-End Dense Visual SLAMVisual geometry

  Show how a foundation model enters a geometric inference system and separate reused components from the intervention.

• URaG: Unified Retrieval and Generation in Multimodal LLMs for Efficient Long Document UnderstandingMultimodal document understanding

  Make retrieval and generation information flow inspectable across long, heterogeneous documents.

• VAEVQ: Enhancing Discrete Visual Tokenization Through Variational ModelingVisual representation

  Explain a discrete representation through its mechanism, reconstruction criteria and generative use.

Training, objectives & optimization (4)

• Modality-Balanced Collaborative Distillation for Multi-Modal Domain GeneralizationMultimodal learning

  Connect modality imbalance to two corrective operations, then test each operation under a named domain protocol.

• ProGMLP: A Progressive Framework for GNN-to-MLP Knowledge Distillation with Efficient Trade-offsGraph learning and distillation

  Separate staged teacher-to-student training from inference stopping and measured runtime.

• Deep (Predictive) Discounted Counterfactual Regret MinimizationGame solving

  Connect expected-target statements to executable neural regret updates without extending a theorem to an unproved implementation guarantee.

• Revisiting (Un)Fairness in Recourse by Minimizing Worst-Case Social BurdenFairness and recourse

  Expose a measurement blind spot before proposing a remedy; keep absolute outcomes beside disparity gaps.

Theory & formal analysis (2)

• Generalization Bounds for Semi-supervised Matrix Completion with Distributional Side InformationStatistical learning theory

  Keep side-information assumptions, theorem scope and rate comparisons together, without requiring decorative experiments.

• On the Learning Dynamics of Two-layer Linear Networks with Label Noise SGDLearning dynamics

  Distinguish an analytically tractable model from the scope of experiments that illustrate its dynamics.

Empirical analysis, diagnosis & auditing (1)

• Why Do Open-Source LLMs Struggle with Data Analysis? A Systematic Empirical StudyLanguage-model analysis

  Organize an empirical investigation around failure questions and controlled comparisons rather than a new-model leaderboard.

Datasets, benchmarks & measurement (1)

• VCapsBench: A Large-scale Fine-grained Benchmark for Video Caption Quality EvaluationVideo evaluation

  Establish annotation provenance and metric denominators before comparing models on a new benchmark.

Efficiency, inference & systems (2)

• CATP: Contextually Adaptive Token Pruning for Efficient and Enhanced Multimodal In-Context LearningMultimodal inference

  Explain where pruning happens and distinguish token, FLOP, memory and end-to-end latency claims.

• DeepProofLog: Efficient Proving in Deep Stochastic Logic ProgramsNeurosymbolic reasoning

  Use symbolic examples to teach a semantic change and distinguish exact execution, approximation and failure.

Domain-grounded methods & physical systems (3)

• Graph-Semantic Guided Learning for Virtual Immunohistochemistry Staining on Consecutive Histology SectionsComputational pathology

  Explain the histology correspondence problem and distinguish image-translation fidelity from demonstrated clinical utility.

• MSAnchor: De Novo Molecular Generation from Mass Spectrometry Data with Anchor-Extended Molecular ScaffoldsMolecular science

  Define the experimental measurement-to-molecule contract and distinguish computational recovery from laboratory validation.

• FoAM: Foresight-Augmented Multi-Task Imitation Policy for Robotic ManipulationRobot manipulation

  Separate anticipated visual information from executed policy behavior and report the setting of success measurements.

Scope, remaining gaps and future additions

Complete the selected deep writing studies first. The remaining official oral inventory stays searchable and explicitly uninspected; it is not a promise to inspect every record in this stage. Main-manuscript writing analysis and selected rendered pages do not constitute proof verification, experimental reproduction or clinical validation.

• No position-paper exemplar has been inspected in this selection; do not infer that the main track contains none.
• Topics are illustrative, not a balanced sample of the 213 oral sessions. Planning, search, multi-agent systems and many other areas remain undercovered.
• Canonical-paper atlas totals describe inspected publisher PDFs. Explicitly separated source-file measurements can include an inspected extended author version and overlapping main text; they are not extra canonical pages or all appendix. This selection cannot estimate conference-wide frequencies of diagrams, tables, section lengths or writing choices.
• Separate supplementary files and extended revisions require their own inspection records. A link, poster or code repository is not evidence that an appendix was read.

Add a paper when it teaches a missing or contrasting explanatory job. Record its selection reason, inspect its actual sources, then update studies, source-bound measurements, synthesis and gallery. Preserve the full inventory and its uninspected status; do not add placeholder studies to satisfy a coverage target.

Conference dates

Each record names the edition: AAAI 2027’s main-track submission deadlines fall in July 2026, while the conference takes place in February 2027. The full meeting period includes ancillary programs, not only technical-paper sessions. Additional venues are labeled by their actual program rather than presented as another full conference.

Dates for AAAI only, newest edition first. Submission dates can fall in the previous calendar year. AoE means Anywhere on Earth (UTC−12); other stated timezones are retained. Reconfirm official instructions before submission or travel.

Compare all conferences on the Research Writing timelines →

AAAI 2027

Main-track abstract

Jul 21, 2026 · 23:59AoE (UTC−12)

Main-track full paper

Jul 28, 2026 · 23:59AoE (UTC−12)

ConferenceMontréal, Québec, Canada

Feb 16–23, 2027Palais des congrès de Montréal

• These are main-technical-track deadlines. Other tracks, including AI Alignment, and individual workshops can follow different schedules; consult their own calls. The 2027 edition received its main-track submissions in July 2026.
• Supplementary material and code were due July 31, 2026 at 23:59 AoE. The two review phases are stages of one main-track submission cycle, not separate submission rounds.
• The full meeting runs February 16–23 at Palais des congrès de Montréal. The registration page specifies tutorials, labs and bridges on February 16–17 and workshops on February 22–23.

Official sources:AAAI-27 main technical track call and timetableAAAI-27 conference and main-track timetableAAAI-27 official venue and program registration dates

Checked Oct 21, 2026.

AAAI 2026

Main-track abstract

Jul 25, 2025 · 23:59AoE (UTC−12)

Main-track full paper

Aug 1, 2025 · 23:59AoE (UTC−12)

Main siteSingapore

Jan 20–27, 2026Singapore EXPO

Offsite workshopsSingapore

Jan 26, 2026Grand Mercure Roxy Hotel

• These are main-technical-track deadlines, not universal deadlines for every AAAI track or workshop. The 2026 edition received its submissions in 2025; supplementary material and code were due August 4, 2025 at 23:59 AoE.
• The full meeting spans January 20–27: pre-conference programs January 20–21, the main technical conference January 22–25, and workshops January 26–27.
• Some January 26 workshops took place at Grand Mercure Roxy Hotel, 50 E Coast Road, Singapore, rather than Singapore EXPO. This is a workshop-only offsite venue, not a second full conference; consult the workshop program for room assignments.

Official sources:AAAI-26 conference, venue and main-track timetableAAAI-26 program overviewAAAI-26 workshop program and offsite venue

Checked Oct 21, 2026.