Machine learning and computing audiences have sat through plenty of talks that open with a leaderboard. The talks that work explain why the problem matters, show enough of the method to believe it, and admit where it breaks.
Quick answer: To present at an AI, data science or computing conference, check which format your paper was accepted for (oral, spotlight, poster or workshop), then turn the paper into a short story: the problem, the method in plain terms, the key result and the limits. Show baselines, error bars and ablations honestly, and prepare for method questions.
Last verified against official sources: October 2026.
This guide covers formats at large machine learning venues, how to turn a paper into a talk, honest reporting of results, demos, code and data, poster sessions, questions and recorded talks.
What Presentation Formats Do AI and Data Science Conferences Use?
Large machine learning venues such as NeurIPS, ICML and ICLR mix oral talks, posters, spotlights and workshop sessions, and the labels mean different things at different conferences. Your acceptance email and the venue’s presenter page decide what applies to you, so read them before you plan anything.
| Format | What it usually means | What to check |
|---|---|---|
| Oral | A short talk on a main stage, followed by a few questions | Total slot length, talking time, slide upload, which laptop you use |
| Spotlight | A designation for selected papers; the exact form differs by venue | Whether it is a short talk, a poster with a sign, or both |
| Poster | You stand beside a printed poster and talk to people who stop by | Board size, orientation, print deadlines, which day and room |
| Workshop talk or poster | A smaller session on a narrower topic, often with different rules from the main conference | Workshop-specific poster size and submission rules |
Some examples show how much venues differ. ICML 2026 schedules each oral for 15 minutes in total and asks speakers to plan for 12 minutes of talking, which leaves time for questions and the change of speakers. The NeurIPS 2025 fact sheet lists 87 orals and 5,290 posters for 5,290 accepted papers, so a talk slot was the exception and a poster was the normal format. The NeurIPS 2025 poster page described spotlight as an honour tag given to selected posters, and it also said accepted papers had to be presented at one of two venues, San Diego or Mexico City. Check the current year’s page, because these rules change between editions.
Industry and practitioner events use other formats. The ODSC AI West 2026 call for speakers lists 30-minute talks, hands-on workshops of 60 minutes or two hours, and a 10-minute AI for Work talk, bans product pitches, and asks speakers to send final slides by a set date in a designated template.
For NeurIPS 2026, the conference runs at three sites, Sydney, Atlanta and Paris, and each accepted paper gets one guaranteed registration for a designated presenting author at an assigned site. The NeurIPS 2026 main track handbook says at least one author of each accepted paper must register for the main conference, and that a Virtual Only Pass is not sufficient. It also says an oral whose authors do not show up can be downgraded to a poster. The NeurIPS blog post on multi-site registration (August 2026) says no changes to the presenting author or poster location are permitted once allocations are finalised, so decide who will travel before you nominate them.
How Do You Turn a Paper Into a Talk?
Do not shrink the paper. Build the talk around four questions: what problem, what idea, what evidence and what limits. A paper is written to be checked line by line, but a talk is heard once, in order, with no way to scroll back.
Start with the problem and why anyone should care, in two or three sentences you could say to a colleague from another subfield. Then give the intuition behind the method before any notation. If your model has a loss function, say what it rewards and what it punishes before you show the equation. Keep one running example, such as a single input and what your method does with it, so the audience can follow the idea without holding the whole architecture in their heads.
Give the key result next, with the comparison that makes it meaningful: better than what, on which data, by how much, and measured how. Finish with the limits, the cases where the method fails or has not been tested. Ending on limits is not a weakness. It tells the audience what to try next and shows you know your own work.
Here is a plan for a 12-minute talk, the speaking time ICML 2026 suggests for a 15-minute oral. Scale it to your own slot.
| Part | Time | What the audience should leave with |
|---|---|---|
| Problem and motivation | About 2 minutes | Why this problem is worth solving |
| Method intuition | About 3 minutes | The core idea, with one example |
| Experiments and key result | About 4 minutes | What you compared and what you found |
| Limits and open questions | About 2 minutes | Where it fails and what comes next |
| Summary | About 1 minute | One sentence they can repeat |
This split is a suggestion, not a rule from any conference. For slide count, the EMBL conference office’s 2018 tips suggest planning roughly one minute per slide, and say a 10-minute talk with many more than 10 slides is likely too long. Put the details that do not fit into backup slides after your final slide, ready for questions. Keep slide text short, and move the longer explanation into your speaker notes or onto a small set of cue cards. If you use presenter view, test it on the same kind of setup you will have on the day, and click through every animation once, because some decks behave differently on a projector than on your laptop screen.
How Do You Present Results Honestly?
Show what you compared against, how many times you ran it, and what did not work. These are the first things an expert listener checks, and a result that holds up under that check is worth more than a bigger number that does not.
Name your baselines and say how you tuned them. If your method got a hyperparameter search and the baseline kept its defaults, say so, or better, give both the same budget. State the dataset, the split and the metric on the slide where the number appears, not only in the paper.
Report variation. The NeurIPS 2026 paper checklist guidelines ask authors to say what the error bars capture (such as the data split, initialisation or random draws), how they were calculated, and whether they show the standard deviation or the standard error of the mean. Even if your venue has no such checklist, a plot with a labelled error bar answers the first question a sceptical listener will ask. If you ran three seeds, say three seeds, rather than implying more.
Show ablations, meaning results with one component removed at a time, so the audience can see which part of your method does the work. If an ablation shows that a component hardly matters, say so. Mention failure cases and negative results too, for example on a slide titled “Where it breaks”.
Be careful with a few common shortcuts:
- Do not tune on the test set, and say clearly which set each number comes from.
- Do not crop a plot’s axis so a small gap looks large.
- Do not call a difference significant without saying how you tested it.
- Do not present a cherry-picked qualitative example as typical. Say how you picked it.
Should You Do a Live Demo?
Only if the demo shows something the slides cannot, and only with a recorded backup ready. A demo that works is the most convincing thing in a talk, and a demo that hangs for two minutes costs you the time you needed for your results.
Record the full demo as a short video before you travel and keep the file on the laptop you will present from, not only in the cloud. Run the live version offline if you can, because conference wifi can be unreliable. Use fixed inputs you have tried before, and prepare one input that shows a failure, so you can answer “does it break?” by showing it rather than describing it.
Check the hardware too. ICML 2026 tells oral speakers to present from their own laptop and to bring any adapter needed for the full-size HDMI cable at the podium, so a demo that needs a special port or dongle should be tested before you walk on stage.
Decide in advance when you switch to the recording. Give the live demo a short fixed time, and if it has not worked by then, say “let me show the recording” and move on without apologising. If the demo takes a long time to run, such as training or a slow model call, start it before the talk and show the result.
How Do You Handle Code, Data and Reproducibility in a Talk?
Say what is released, where, and under which licence, and keep the answer to one slide. Several large machine learning venues ask about reproducibility, and someone in the audience may ask whether they can run your code.
NeurIPS does not require you to release code. Its 2026 paper checklist treats “no, because the code is proprietary” as an acceptable answer, and its code submission policy strongly encourages submitting code that produces the main result, anonymised for review. The ICLR 2026 author guide also encourages code and says reviewers are encouraged, but not required, to look at supplementary material. Rules differ by venue and year, so read the one that applies to you. In the ACM world, the artifact review and badging policy defines separate badges for artifacts that are available, artifacts that reviewers evaluated (functional or reusable), and results that were reproduced or replicated. Those badges apply only where a conference or journal runs that process, so do not mention a badge unless your paper actually has it.
A short reproducibility slide can cover the following:
- Where the code and trained models are, and whether they are public yet.
- Which datasets you used, their licences and any consent or privacy limits on them, such as patient privacy rules for clinical data.
- The compute you used, in plain terms, such as the type of hardware and rough run time.
- The random seeds, library versions and anything else needed to rerun the main result.
If you cannot share data or code, say why (privacy, licence, commercial restriction) and what you offer instead, such as a smaller public dataset or a detailed description of the setup. Do not promise a release date you cannot keep.
How Do You Run a Poster Session at a Large Venue?
Prepare a 30-second version, a 2-minute version and a longer walk-through, because people at a poster session arrive at different depths. Visitors decide quickly whether to stay, so lead with the problem and the main result.
Check the size rules early, because printing deadlines come before travel. For ICLR 2026, the poster page gives a maximum of 1.90 m by 0.90 m (76.4 in by 37.4 in) in a horizontal layout for main conference posters, and a maximum of 61 cm by 91 cm (24 in by 36 in) in a vertical layout for workshop posters. ICML 2026 suggests 36 by 48, 36 by 60 or 36 by 72 inches (height by width) for the main conference, caps workshop posters at 36 inches high by 24 wide in portrait format, and set a print order deadline of 15 June 2026 for its poster printing service. Do not reuse your main poster for a workshop without checking its limit.
Design the poster to be read from a short distance:
- Put the title, the one-sentence claim and your contact details at the top.
- Give the main result a large figure with a clear caption, and keep paragraphs short.
- Add a code or paper link, such as a QR code, only if the link works and is public.
- Leave out equations you will not discuss.
When someone walks up, ask what they work on. That lets you aim your explanation at their background instead of repeating one pitch to everyone. Keep a notebook for names and suggestions, and bring water, because you will talk for hours.
How Do You Answer Questions About Your Methods?
Answer the question that was asked, briefly, and say “I do not know” when that is true. A confident wrong answer costs you more credibility than an honest gap, and you can offer to follow up.
Before the talk, write down the hardest questions a reviewer might ask and prepare a backup slide for each. Typical ones include why you chose this baseline, whether the result holds on another dataset, how sensitive it is to hyperparameters, what the compute cost is, and whether the test data could overlap with the training data. If you can point to a backup slide, the answer feels solid even before you speak.
In the room, repeat or rephrase the question so everyone hears it, especially in a large hall. If the question contains a criticism, restate the strongest version of it before you answer. If you need more time than the session allows, offer to continue in the corridor or by email.
What If the Venue Wants a Recorded or Virtual Talk?
Follow the venue’s instructions for that year, because the rules vary. For example, the ICLR 2026 poster page calls video recording “highly recommended”, and its author guide says authors who registered to attend record a 5-minute summary video that is hosted for asynchronous viewing. Check whether your event has changed this.
If you do record, treat it as its own talk, not a screen capture of the live one. Speak a little slower, avoid phrases like “as you can see here” unless your cursor is visible, and keep the file within the length the venue allows. Check the audio before you upload.
From our experience at MS International: For most applicants, the route to our in-person conferences in the Greater Toronto Area is the presenting-author track: submit work for review and register after acceptance. We accept abstracts, full papers, case studies, review papers and professional presentations, so computing and data science work can go in as any of these, with full papers normally four to twelve pages in APA format and the work original and not published or under review elsewhere. Review is double-blind, and decisions are usually returned within about 10 working days, although complex papers or revision rounds can take longer. Keynote and invited-speaker roles are arranged separately with the organizing team and are not obtained by buying an author ticket. Speakers cover their own travel and personal costs unless agreed in writing, and accepted papers may be considered for conference proceedings where applicable. Presentation lengths may vary, so the speaker confirmation and final programme are the official sources.
FAQ: Presenting at an AI, Data Science or Computing Conference
Spotlight or oral, live demo or recorded, code released or kept private: AI and computing presenters weigh these alongside poster size and virtual attendance at NeurIPS 2026.
How long is an oral talk at an AI conference?
It is often around 15 minutes in total, but it varies by venue. ICML 2026 schedules orals for 15 minutes and asks speakers to talk for 12, leaving time for questions. ODSC AI West 2026 lists 30-minute talks. Check your acceptance email for your own slot.
What is the difference between a spotlight and an oral?
It depends on the venue. At NeurIPS 2025 and ICML 2026, spotlight papers were marked with a tag or sign at the poster session rather than given a separate stage talk. Other venues may differ, so read the presenter page for your event.
What size should my poster be?
Use the size given by the venue for that year. ICLR 2026 allowed up to 1.90 m by 0.90 m for main posters, and workshop posters at ICLR 2026 and ICML 2026 were smaller and portrait. Sizes change by event, so check before printing.
Do I have to release my code?
No, not at NeurIPS: its 2026 checklist accepts “no” with a reason such as proprietary code, though it strongly encourages submitting code that produces the main result. ICLR 2026 also encourages code. If you cannot share it, say why and what you offer instead.
Should I do a live demo in my talk?
Only with a recorded backup ready. Keep the recording on your own laptop, decide in advance when you will switch to it, and use inputs you have already tried.
Can I present virtually at NeurIPS 2026?
No, a virtual-only registration does not meet the NeurIPS 2026 rule for accepted papers. At least one author must register for the main conference and present at the assigned site. If serious circumstances stop all authors attending, ask the organisers whether someone else can present.
Who pays for my travel as a speaker at MS International?
You do. Speakers cover their own travel, accommodation, visa, transportation and personal expenses, and participation creates no entitlement to an honorarium, speaker fee or travel reimbursement unless MS International has agreed otherwise in writing.
Conclusion
A good talk at a computing or data science conference is short on claims and long on evidence: one clear problem, an idea people can repeat, results with baselines and error bars, and an honest list of limits. Check the format and rules for your event first, prepare a backup for anything that can fail, and rehearse your answers to the hardest questions.
