ICLR — writing guide

Make the new knowledge easy to inspect: explain the problem, what changes, why the change matters, which evidence tests it, and where the conclusion stops. That is a more useful writing target than imitating the appearance of an oral paper.

This overview synthesizes the ICLR 2026 oral-paper studies. It offers editorial lessons, not an official consensus, a unique style shared by every ICLR paper, or a formula for acceptance. The reviewer guidance asks about the problem, motivation, support for claims and contribution; value need not mean a new leaderboard record.

Read the short synthesis · Choose detailed material · Use a drafting checklist

Learn from the complete oral corpus

The 223-paper writing atlas compares figure/table counts, main and appendix spans, and category-level distributions. Its searchable ledger leads to a section-by-section layout record for every paper, with source pages and extraction limitations. These are measured or estimated observations, not suggested submission quotas.

Corpus snapshot: extraction-assisted medians are 5 main figures and 3 main tables. The median complete PDF is 30 physical pages, including references and any attached appendix—not a submission-length recommendation. Some theoretical papers use neither figures nor tables; the full distributions and category differences matter more than matching a median.

Corpus statistics and every paper’s layoutSee distributions across all 223 PDFs; compare contribution types; search figures, tables, section lengths and appendix records.

What to prepare for your type of contributionSix category syntheses: architecture, learning/optimization, systems, theory, empirical studies and benchmarks. Each has an outline and conditional figure/table/algorithm package.

See the visual forms directlyTwelve inline examples, including attributed result tables and original diagrams, curves, heatmaps and qualitative layouts—with caption checklists.

Writing in brief

Claims and wording: precise rather than promotional

State what changes relative to the closest work, using concrete subjects and verbs. Keep “we observe,” “we hypothesize” and “we prove under these assumptions” distinct. Replace unsupported “universal,” “first” or “significant” language with the tested setting and evidence. Reuse the same terminology in the abstract, contributions, figures and conclusion.

For an example of connecting obstacles to interventions, see Mamba-3’s framing study. For sentence-level practice, use the style guide.

Structure: follow the argument, not a fixed list of headings

A useful starting sequence is problem → gap → idea → mechanism → evidence → limits. The abstract should give the reader a compact version of that argument; the introduction establishes why it matters. Introduce definitions before relying on them, and organize experiments around questions that test the contribution. Theory, datasets and systems need different outlines.

The InfoNCE writing study illustrates the importance of keeping assumptions and guarantees together. See contribution-specific outlines for alternatives, not mandatory headings.

Visualization: make the changed computation visible

Give each figure one explanatory job. A mechanism diagram should distinguish familiar components from the intervention, label ambiguous arrows, and separate training-only signals from inference inputs. A caption explains the setting, symbols and supported observation. Keep fonts readable at publication size; use shapes, labels or line styles as well as color.

See TD-JEPA’s visual explanation and the original architecture/caption blueprints. For a performance plot, show the relevant quality–resource trade-off rather than a speed claim without its denominator.

Evidence and tables: let readers audit the comparison

Put the dataset, split, metric direction, resource budget and uncertainty near the numbers. Separate main comparisons, controlled ablations and deployment costs. An ablation should challenge the proposed explanation, not simply remove arbitrary modules. Retain settings where a baseline wins and distinguish measured results from fitted or extrapolated trends.

The FlashVID study shows why token reduction and end-to-end latency are different claims. The comparison-pattern guide provides further examples.

Appendix and reproducibility: build an evidence trail

Use stable references from each main claim to its proof, full results, protocol or configuration. Keep essential assumptions and the defining method intelligible in the main text. Use the appendix for verification detail, not to hide failures or the only explanation of how the method works.

See OpenThoughts’ appendix study and the claim-indexed appendix plan.

Format and polish: check the exact edition

Keep notation, model names, figure labels and metric units consistent. Review the compiled PDF for legibility, broken equations, table alignment and reference accuracy. Apply the official template for the submission year and phase; an accepted camera-ready PDF does not establish the original submission rules. The 2026 requirements page retains the verified limits, template and source conflicts rather than treating them as timeless ICLR rules.

Explore ICLR 2026

223 inspected writing studies out of 223 oral records in the dated program snapshot. Choose a writing lesson, an individual paper, or the historical submission requirements.

Writing patterns and worked examplesStart here to learn a technique: framing, visual grammar, comparisons, efficiency claims, appendices and calibrated wording.

Search the oral-paper writing studiesFind a paper, topic or writing technique; inspect its organization, figures, tables, source locations and limitations.

2026 rules, formatting and official templateCheck phase-specific requirements, anonymity, reproducibility, disclosure and documented disagreements between official sources.

The synthesis is based on this edition, not a comparison with rejected papers or other conferences. Additional editions will keep their own sources and requirements; shared lessons should be revised when the evidence changes.

Edition archive

The ICLR dates below list submission deadlines and event locations independently of these inspected writing editions.

ICLR 2026

223 / 223 inventoried oral papers have inspected writing studies. Snapshot: 2026-09-19.

Corpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

Each edition retains its own sources, coverage and requirements. The synthesis above remains tied to its named edition.

Conference dates

ICLR submission deadlines, event periods and locations are listed below, independently of the inspected writing editions. Unannounced dates are not estimated.

Dates for ICLR 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 →

ICLR 2027

Abstract

Sep 18, 2026AoE (UTC−12) · clock time not separately specified

Full paper

Sep 25, 2026 · 23:59AoE (UTC−12)

ConferenceSan Francisco, California, USA

Apr 26–30, 2027

• The 2027 edition had its submission deadlines in September 2026. The abstract is listed as an AoE date without a separate clock time; the author guide explicitly gives 11:59 PM AoE for the full paper.
• The displayed event period includes the main conference on April 26–28 and workshops on April 29–30.
• The homepage lists California and says venue information will be posted. The workshop call identifies San Francisco; its workshop-year text contains a 2026 typo. The period here follows the 2027 homepage and Dates schedule. The venue is left unconfirmed rather than inferred from an image caption.

Official sources:ICLR 2027 call for papersICLR 2027 author guidelinesICLR 2027 datesICLR 2027 conferenceICLR 2027 call for workshops

Checked Oct 21, 2026.

ICLR 2026

Abstract

Sep 19, 2025 · 23:59AoE (UTC−12)

Full paper

Sep 24, 2025 · 23:59AoE (UTC−12)

ConferenceRio de Janeiro, Brazil

Apr 23–27, 2026Riocentro Convention and Event Center

• The 2026 edition had its submission deadlines in September 2025.
• The displayed event period includes the main conference on April 23–25 and workshops on April 26–27.
• The Dates page retains a generic location-and-dates-to-be-announced sentence; the explicit dated schedule and conference homepage confirm the period and venue.

Official sources:ICLR 2026 call for papersICLR 2026 datesICLR 2026 conference and venue

Checked Oct 21, 2026.

Apply it to your draft

1. Write one bounded contribution sentence and identify its closest comparison.
2. Sketch the main figure so a reader can explain the changed mechanism.
3. Match each claim to a test, table or theorem, including its assumptions.
4. Link each result to the protocol or appendix detail needed to check it.
5. Recheck the title, abstract and conclusion against the final evidence and the exact edition’s rules.

Use the argument and evidence worksheet to work through these steps, then ask an independent reader to explain the claim, mechanism, evidence and limitation without your help.