NeurIPS 2026
68 / 106 inventoried oral papers have inspected writing studies. Snapshot: 2026-10-02.
NeurIPS — writing guide
Make the new knowledge and its scope inspectable. State the question, distinguish the intervention from the closest alternative, explain the reasoning and show evidence that can test it. A theorem, a benchmark, an empirical finding and a deployed system require different arguments.
The latest source is the NeurIPS 2026 accepted-oral program snapshot, frozen on 2 October 2026 at 09:33:46 UTC. It contains 106 distinct papers with explicit oral decisions, including 100 with Oral event entries and six recorded only as posters so far. The records span 95 Main Track and 11 Evaluations & Datasets papers. The program is live and the conference has not yet taken place. These are not final proceedings statistics.
Choose the level of detail
Section titled “Choose the level of detail”| You want to… | Start here |
|---|---|
| See precisely which manuscripts are inspected or unavailable | Complete accepted-oral inventory and source status |
| Compare figure/table counts and section spans of available versions | Measured-subset atlas |
| Plan an architecture, objective, theory, analysis, resource or system paper | Contribution-specific writing plans |
| See an ablation table, calibration plot or explanation diagram | Inline visual gallery |
| Learn argument and evidence patterns | Writing patterns |
| Check main-track format, checklist and submission phases | 2026 main-track rules |
Writing in brief
Section titled “Writing in brief”Claims and wording: keep the quantifiers visible
Section titled “Claims and wording: keep the quantifiers visible”Name the setting, intervention and supported conclusion. Distinguish an exact identity from an approximation, a guarantee under assumptions from a measured trend, and a diagnostic subset from a population-level result. Avoid turning “works in these settings” into “works universally.” Keep the same terms in equations, captions and contribution statements.
The Gaussian moment-matching study distinguishes exact layerwise moments from approximate deep-network output distributions. The SyncWorld study retains the oracle-headroom condition for a selected policy-improvement table. These qualifications are part of the result itself.
Structure: align the outline with the burden of proof
Section titled “Structure: align the outline with the burden of proof”For an empirical method, a useful sequence is question → failure or mismatch → mechanism → discriminating test → limits. For theory, introduce the objects and assumptions before the statement and proof roadmap. A benchmark must explain the intended construct and evaluator before ranking systems. A systems paper must define the workload and quality contract before claiming speed.
An early comparison of assumptions can make a theoretical gap more precise than several paragraphs of broad novelty language. An architecture paper can instead use a worked input/output example. The type guides explain which displays to prepare and when pseudocode, a theorem or a diagram would be unnecessary.
Visualization: distinguish the object, intervention and observation
Section titled “Visualization: distinguish the object, intervention and observation”An architecture overview should show the actual representation and changed operation. Name frozen versus trainable parts, conditioning versus supervision, and what remains at inference. In BEAST3D, the diagram separates per-view features, trainable multi-view fusion, splats and held-out-view targets. In Surflo, the shared latent is distinguished from variable-resolution decoding.
Use plots to answer a question: variation across seeds, response to an intervention, an accuracy–cost trade-off, or calibration relative to a nominal target. Use a table when exact lookup matters. Define every visual unit—one point may be a neuron, task, run or model, with very different implications.
Evidence and tables: test the explanation, not only the final score
Section titled “Evidence and tables: test the explanation, not only the final score”Keep data, supervision, model capacity, tuning budget, inference information and resource accounting close to the comparison. Distinguish a matched control from broader leaderboard context. The PhaseLoRA table example separates extra capacity, arbitrary descriptors and scalar modulation as alternative explanations.
State what uncertainty means. A standard deviation across seeds is not a confidence interval for a paired difference. A calibration target is not a quantity to maximize indefinitely. Preserve non-wins: the equivariant pruning study keeps benchmark metrics where the baseline is better and distinguishes training cost from deployment cost.
Appendix and checklist: make the evidence traceable
Section titled “Appendix and checklist: make the evidence traceable”Give readers named routes to assumptions, complete proofs, task protocols, prompts, configurations and full results. Keep the central argument intelligible without those details. Long appendices in author reports demonstrate navigational choices, not recommended appendix lengths.
For the 2026 Main Track, the paper, references, textual appendices and mandatory checklist belong in one PDF. Checklist answers should be truthful and point to actual evidence. This is a reporting obligation, not a request to mark every item “yes.” Evaluations & Datasets has separate track instructions; the main-track rules must not be imposed on it automatically.
Format and polish: verify the track and phase
Section titled “Format and polish: verify the track and phase”The checked Main Track rules allow nine content pages initially and ten after acceptance, excluding references, technical appendices and the checklist. Use the official single-column, 10-point template unchanged. Numeric or author-year citations are permitted when consistent; figure captions go below and table captions above.
Several available author manuscripts use other formats or are long technical reports. Their layout measurements do not demonstrate NeurIPS submission compliance. The rules page separates official requirements from those observations. Future editions must be checked independently.
Edition archive
Section titled “Edition archive”Each edition retains its own sources, coverage and requirements. The synthesis above remains tied to its named snapshot.
Conference dates
Section titled “Conference dates”NeurIPS submission deadlines, event periods and locations are listed below, independently of the inspected writing editions. Unannounced dates are not estimated.
Dates for NeurIPS 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 →
NeurIPS 2027
- Submission
- Not announced
- ConferenceEurope — city and country unannounced
- Dates not announced
- Only the region is announced. Submission deadlines, event dates, city, country, and venue were not announced in the official sources checked; no dates are extrapolated from 2026.
Official sources:NeurIPS future meetings
Checked Oct 21, 2026.
NeurIPS 2026
- Abstract
- AoE (UTC−12) · clock time not separately specified
- Full paper
- AoE (UTC−12) · clock time not separately specified
- Main siteSydney, Australia
- Dec 6–12, 2026International Convention Centre Sydney
- Satellite siteAtlanta, Georgia, USA
- Dec 9–13, 2026Georgia World Congress Center
- Satellite siteParis, France
- Dec 9–13, 2026Palais des Congrès de Paris; Sorbonne University (Jussieu) for selected workshops on December 12–13
- These are main-track deadlines; the full-paper deadline includes all supplementary materials. The official sources give AoE dates without a separate numerical clock time.
- Sydney is the main site: expo December 6, tutorials December 7, main conference December 8–10, and workshops/competitions December 11–12.
- Atlanta and Paris are official satellite sites, both December 9–13. Atlanta has main sessions/tutorials December 9 and main sessions December 10–11; Paris has main sessions/tutorials December 9–11. Both hold workshops December 12–13. Do not treat the three locations as one continuous Sydney event.
- Paris uses two venues: the Palais des Congrès for the main meeting and workshops, and Sorbonne University (Jussieu) for selected December 12–13 workshops. Check the workshop schedule for the correct building; Paris remains one satellite city, not two separate conference editions.
Official sources:NeurIPS 2026 main-track call for papersNeurIPS 2026 dates and three-site scheduleNeurIPS 2026 conferenceNeurIPS 2026 official venue informationNeurIPS 2026 attendee support — Paris workshop venues and dates
Checked Oct 21, 2026.
Apply it to a draft
Section titled “Apply it to a draft”- Write the claim with its setting, assumptions and comparison conditions.
- Choose a contribution type and identify the decisive evidence or proof obligation.
- Explain the changed computation or mathematical object using a worked example when useful.
- Pair task results with controls that test the proposed mechanism and report uncertainty correctly.
- Make limitations and appendix destinations explicit, then complete the correct track’s reporting requirements.
- Inspect the compiled PDF, not just LaTeX source, for readability and rule compliance.
The argument worksheet supports this process. The collection offers editorial lessons, not an official consensus, complete source coverage, or a formula for acceptance.