How to turn a research paper into a short conference presentation

Disclosure: NOXI publishes this guide and is one of the tools it mentions; the method works in any presentation software.

Compressing every section of the paper into fewer words usually produces a talk that feels rushed and still hides its contribution: the literature review consumes the opening, the methods become an unreadable diagram, and the main result appears just as the session chair signals that time is up. This guide works through an explicitly fictional sensor study, with invented data and reproducible arithmetic, to show how to choose a claim, keep the evidence that makes it defensible and build a ten-minute talk without turning a small study into a sweeping promise.

What a short conference talk has to do

A conference talk has one job: to help other researchers understand a single contribution well enough to judge and discuss it. It is not a shorter copy of the paper, and equal coverage of every section is rarely the right use of the time.

The most useful preparation question is: "What should another researcher understand well enough to discuss after this talk?" The answer might be a surprising result, a methodological contribution or a limitation that changes how the field interprets previous work. Whichever it is, the talk needs four elements around it:

• A clear research question that names the unresolved issue.
• Enough method to judge the answer, chosen by what affects belief in the claim, not by what took the most work.
• Visible evidence, in a form the audience can read while you explain it.
• An honest boundary that says where the conclusion stops.

Start with those four elements, then decide which parts of the paper help the audience understand them. Everything else has another home: backup slides, the paper or the supplement.

Establish the presentation contract first

The organizer's instructions decide how long you speak, what the file must contain and how questions are handled, so read them before planning a single slide. Record the speaking allowance separately from the total slot, and check whether introductions, setup or questions are included.

Two 2026 events show how much this varies. CHI 2026 allocated on-site paper talks ten minutes of presentation plus two minutes of questions. ETRA 2026 allocated twelve minutes plus three minutes of questions for full papers and selected short papers. These were the requirements for those events, both checked on 7 October 2026; they are not a rule for every ACM conference.

The preparation sheet

Capture these items in one place before drafting. Each one changes the content, not only the logistics.

How the format changes the talk

None of these formats needs a slide count fixed before the argument exists. If the timing is unclear, ask the session chair.

Choose the contribution before choosing the slides

Decide what the talk claims before deciding what it shows. Write three sentences privately: the question, the answer the evidence supports, and the boundary beyond which a conclusion would be unreasonable.

The phrase "single-batch bench test" is part of the finding's scope, not a disclaimer to hide on the final slide. The fictional paper could also contain calibration details, several supplementary checks and a long account of competing approaches. The talk's contribution is the held-out test: the correction was fixed before the twelve evaluation units were measured. Keep that development/evaluation separation visible, because it determines how the audience should interpret the result.

What counts as sufficient evidence depends on the kind of research

Choose the method details that affect belief in your particular claim. The table applies that rule to common types of contribution.

Two cautions: a qualitative talk should not suggest that three quotations represent an entire population, and better benchmark performance does not by itself establish better performance in every deployment. The most useful retained method detail is the one that lets the audience judge the evidence your paper actually supplies.

The worked example: a fictional sensor study

The example is a complete small study, so every editing decision in this guide can be checked against the full record.

The researchers developed a linear correction using six development units, then froze its parameters. Twelve different units, randomly selected from one production batch, formed the evaluation sample. No evaluation unit was used to tune the correction, and none was excluded after measurement.

Each evaluation unit was measured at six stable chamber temperatures (15, 20, 25, 30, 35 and 40°C), with three settled readings at each temperature. The factory calculation and the frozen correction were both applied to the same captured readings, then compared with the same laboratory reference.

For each unit, mean absolute error (MAE) was the average of eighteen absolute differences from that reference; lower values mean closer agreement with the reference over this test's temperature range. The device was the unit of comparison. Twelve units × six temperatures × three readings gives 216 captured readings, but the repeated readings estimate each device's error; they do not create 216 independent devices. Applying two calculations to the same readings does not produce 432 independent observations either.

All twelve unit-level results

This table is useful for checking the example. In the real talk it belongs in the backup material, not the opening.

The mean reduction is 0.17°C, approximately 35% of the baseline mean. "35% lower mean absolute error" describes that calculation; "35% more accurate" is less precise, because it names neither the metric nor the denominator. Individual reductions range from 0.08°C to 0.28°C. Every unit improves in this constructed dataset, but that observation does not establish that every future unit will improve.

The uncertainty around the mean

An illustrative paired t interval for the mean reduction is approximately 0.14°C to 0.20°C at the 95% confidence level. It is calculated from the twelve device-level differences: a sample standard deviation of about 0.0471°C, a standard error of about 0.0136°C and a critical t value of 2.201 with eleven degrees of freedom. The model assumes independent, approximately normally distributed unit-level differences. It does not incorporate shared reference bias or establish performance in a wider operating environment.

The NIST/SEMATECH engineering statistics handbook describes analyzing naturally paired measurements through their differences (NIST: paired observations) and explains the repeated-sampling interpretation of confidence intervals (NIST: confidence limits for the mean). A 95% interval is not the range of individual device improvements, and it is not a 95% probability that a future device will improve by that amount.

The result is promising within its scope. It still leaves important questions open: how performance changes between batches, whether the correction remains useful after aging, and whether reference or chamber effects alter the estimate. Those questions belong in the talk, because they determine the next experiment.

Decide what the audience needs now, later or elsewhere

Give every piece of paper material a destination: the main talk for what listeners need to follow and assess the contribution, backup slides for likely questions, and the paper and supplement for the full record. Editing then becomes sorting rather than a series of painful deletions.

The equation decision is conditional. A methods conference may need the equation early. A results-focused session may be better served by a diagram showing what the correction receives and changes. Removing an equation from the main sequence should not remove an essential assumption with it.

Redesign the evidence for the room

A paper figure is read beside a caption and a detailed argument; a slide figure has to become understandable while you explain it. Redraw it around the one message it supports, and keep the unit, sample size, pairing and comparator readable.

MIT's Civil and Environmental Engineering Communication Lab illustrates the distinction with original figure-redesign examples, including replotting selected data and identifying the adaptations made (MIT CEE Communication Lab: redesigning existing figures for slides, published 15 January 2024).

Imagine the fictional paper's main figure has four panels: development fit, errors at each temperature, unit-level evaluation and residual diagnostics. Copying all four onto a slide would require the audience to learn several graphical conventions before seeing the contribution.

The main result as a paired dot plot

For the main result, use a paired dot plot. Put factory MAE on the left and corrected MAE on the right, with a line connecting the two values for each unit. Use one vertical scale labeled "Mean absolute error, °C", running from zero to about 0.65°C so that every value in this example fits, and say that each line represents one test device. Direct labels can identify the two group means without making them look like additional observations.

Before revealing the points, teach the audience how to read them: "Each device has two error estimates from the same readings. A line moving down means that the correction reduced error." Then show the result. The audience now knows what to inspect.

A second, smaller panel can show the mean paired reduction and its interval, labeled as an interval for the mean. Do not present that panel as independent confirmation: it summarizes the same twelve devices.

Figure redesign checklist

Redrawing a chart does not turn another team's results into your own data. For crediting data and figures on slides, see our guide to presentations with research and citations.

A 10-minute conference presentation, slide by slide

For a ten-minute speaking allowance, nine main slides plus three backup slides form a workable structure, with the evidence slides getting the most time and the title and closing slides the least. The template below is an editorial model, not a mandatory format; its times include spoken explanation and ordinary transitions.

The template

Write every title as a full-sentence claim that stays true when read alone. "Results" tells the audience nothing; "Mean absolute error fell from 0.48°C to 0.31°C" tells them what to look for, and someone skimming the PDF later still gets the finding.

The template filled in for the sensor study

1. 0:00–0:40 · "Can a fixed correction improve unseen sensors?" Give the question, presenter and team, plus the result and its laboratory scope in one sentence, without every number. Opening line: "We tested whether a correction developed on separate units would reduce error on twelve sensors from a different evaluation sample."
2. 0:40–1:30 · "A correction needs an evaluation beyond its development units" A fitted correction can look useful on the units used to develop it; the question is how it performs on held-out units. Show a simple development-to-evaluation diagram, and avoid claiming that earlier work never validated corrections: the study supplies no literature evidence for that.
3. 1:30–2:10 · "The correction was frozen before the twelve-unit test" Show six development units feeding parameter selection, then a clear separation before twelve evaluation units, all of which remained in the analysis. This short slide establishes the condition that makes the next result meaningful.
4. 2:10–3:20 · "Each device contributed one paired error comparison" Show the temperature range, three readings at each point and the eighteen-reading summary per device. Define MAE in plain language, explain why twelve device summaries are compared rather than 216 readings, and end by teaching the audience how to read the paired plot.
5. 3:20–5:10 · "Mean absolute error fell from 0.48°C to 0.31°C" Show the twelve paired values. State the 0.17°C mean reduction, then the variation: individual reductions range from 0.08°C to 0.28°C. Pause for the figure, with no paragraph competing with your explanation. This is the center of the talk and gets the most time.
6. 5:10–6:10 · "The mean improvement has uncertainty and a defined scope" Show the illustrative 0.14–0.20°C interval with its label. Say what it estimates and what it leaves out, and distinguish a mean reduction from a guaranteed improvement for an individual device. The calculation stays in backup.
7. 6:10–7:30 · "This test does not cover humidity, drift or other batches" Connect each limitation to an interpretation: one batch limits manufacturing generalization, stable chamber conditions do not establish behavior in a changing environment, and short-term measurements do not establish durability.
8. 7:30–8:40 · "The next evaluation must test the missing conditions" Propose a follow-up with independent batches, operating conditions relevant to the intended use and repeated measurement over time, labeled as proposed checks, not findings. Say what would change the conclusion: a batch effect or a deteriorating improvement would narrow where the correction is useful.
9. 8:40–9:30 · "The correction reduced lab error; field performance remains untested" Restate the result and its boundary, thank collaborators and show where the full paper and measurement record can be found. End with a question worth discussing, such as which deployment condition the next test should prioritize.

Keep the final thirty seconds free. If the venue gives you twelve speaking minutes, spend the extra time where explanation helps most, such as the measurement assumptions or the interpretation of device-level variation. Adding unrelated supplementary results simply because two minutes exist can weaken the narrative.

How many slides for 5, 12, 15 or 20 minutes

Let the argument set the slide count and a timed rehearsal confirm it. These planning ranges keep the same logic at other lengths; they are editorial guidance, not conference rules.

Our slides-per-minute table gives the general arithmetic for 5- to 60-minute talks. For a research talk, weight it by evidence: slides that show data need more time than the title or closing slide.

Write the transitions before polishing the slides

Spoken transitions carry the reasoning from one slide to the next, so draft them before working on layout. If you cannot explain why one slide leads to the next, the sequence needs another cut or a better bridge.

For the sensor study, three transitions do most of the work:

• "A good fit on development units would not answer our question, so we fixed the correction and evaluated different units."
• "These repeated readings summarize each device's performance; now we can compare the two calculations within the same device."
• "That result answers the bench-test question, but the test design leaves three deployment questions open."

Sentences like these keep the talk from becoming a tour of headings labeled Introduction, Methods, Results and Discussion. A reusable pattern is "[What we had] would not answer [the question], so we [did this]." Write one for every slide boundary; a boundary where no sentence comes to mind is where the logic is missing.

How much methodology to include

Include enough method for a specialist to audit the path from measurement to claim: how units were selected, how development and evaluation were separated, how the endpoint was constructed, what was compared and which assumptions matter. Other settings can stay in backup or the paper.

A specialist question often targets exactly the detail a presenter is tempted to shorten. For the sensor study, a listener should be able to answer five questions after the main talk:

1. What was measured?
2. What changed between the two calculations?
3. What stayed paired?
4. What counted as one unit?
5. Which operating conditions were tested?

They do not need to remember every instrument setting. If a setting is essential to validity, its implication belongs in the main explanation; the exact value can remain in backup or the paper.

For the sensor study, prepare three:

A backup section organized as a copy of the entire supplement is difficult to navigate while an audience waits.

How to present limitations and uncertainty

Present each limitation with the interpretation it restricts and the test that would resolve it, and put the scope into the claim itself rather than onto a final slide. A generic list of limitations whose consequences remain unexplained does not help anyone judge the result.

Wording that matches the evidence

Many overclaims enter through wording rather than through the data. Compare each claim-bearing title and sentence with what the analysis supports.

The same discipline applies to ranges and estimates outside research; our guide to presenting forecasts without false precision covers it for planning figures.

Prepare answers that separate evidence from interpretation

Prepare each likely question in three parts: the short answer, the record that supports it and anything that remains unresolved. That preparation makes a response more useful than improvising a broader claim under pressure.

For any paper, the same categories recur: sample size and selection, independence and the unit of analysis, what the uncertainty estimate covers, generalization beyond the tested conditions, practical significance against a specification or need, and comparison with the strongest alternative approach. Prepare at least one question in each category, and check that a backup slide or a sentence in the talk answers it.

When a question reveals a missing analysis, say so directly. Explain whether the necessary data exist and what analysis would answer the question. A precise "we have not tested that condition" preserves the distinction between findings and plans.

Rehearse the argument, then check the delivered file

Rehearse at least twice with different reviewers, then inspect the exact file you will present on the expected display.

During the first timed rehearsal, note when the audience reaches the main result; in the nine-slide plan it arrives at 3:20. If it arrives later, inspect the background and method explanations before speaking faster, and remove information whose main purpose is to demonstrate how much work was done.

For the second rehearsal, ask a colleague unfamiliar with the paper to identify the claim, the comparator and the most important limitation. Ask a specialist where the method explanation leaves a validity question unanswered. These are different reviews, and both are useful. The IEEE Author Center also recommends recording a practice talk and leaving time for audience questions, and Christopher Lortie's ten simple rules for short and swift presentations (PLOS Computational Biology, 2017) make the related point that a short talk needs deliberately reduced scope and practice.

Final file checklist

For a fuller access check, use our presentation accessibility checklist.

Use AI after the scientific decisions are made

An AI presentation tool can turn an approved research brief into a first layout and suggest alternative wording, while the researcher still chooses the contribution, approves the data and judges whether each visual supports the conclusion. Make those decisions first, then give the tool the result.

The most useful input is a concise, approved brief rather than the whole paper: the timing contract, the audience, the central claim, the retained method details, the approved figures and data table, and the overclaims to avoid. A tool given only the complete PDF cannot know which pages carry the contribution, so the draft is likely to follow the paper's proportions, which is the problem this guide started with. Converting a whole document suits other jobs, such as a reading summary; our PDF-to-presentation guide covers them. If the conference or your lab provides a slide template, include it as well; our comparison of AI presentations in your own template explains the options. Prompt 1 below is a complete brief for the sensor study.

Review the output against the paper, not against how polished it looks:

• Numbers and units: every value, unit, denominator and sample size matches the paper.
• Error bars and intervals: the meaning of each one is stated and unchanged.
• Claim-bearing titles: no title claims more than the evidence, and the scope words survive.
• Diagrams: a generated diagram that implies the correction was trained on evaluation units needs correction, however attractive the layout.
• Figures: approved scientific figures stay intact unless you can verify a redesign against the underlying data; when a chart must be redrawn, supply the numbers rather than a picture of the old chart.

For unpublished or restricted material, use only inputs authorized by your institution and collaborators. Our guide to preparing confidential material for an AI presentation tool covers what to remove first.

3 AI prompts for a conference talk

Written for NOXI after attaching your approved brief, figures and data table; they work in any AI presentation tool. Replace the brackets.

How to build the conference talk in NOXI

NOXI drafts the talk from your approved brief in your conference or lab template. It is free, with no card and no watermark; uploading files needs a free account.

Frequently asked questions

How do you turn a research paper into a conference presentation?

Start from the contribution, not the paper's sections. Write three sentences: the question, what the evidence supports and what it does not. Keep only the method a listener needs to judge that answer, redraw the main figure for projection, give the result the most time, and move equations, full tables and extra checks to backup slides. Then rehearse with a timer.

How long is a conference presentation?

As long as the organizer specifies, and the speaking time is often shorter than the slot. In 2026, CHI gave on-site paper talks 10 minutes plus 2 minutes of questions, and ETRA gave full papers and selected short papers 12 minutes plus 3. Read your own event's instructions; one conference's timing is not a rule for others.

How many slides should a 10-minute conference presentation have?

About eight to ten main slides, plus backup slides. The worked example in this guide uses nine: question, gap, design, how to read the evidence, main result, uncertainty, limitations, next test and take-home. Slides do not take equal time: the main result gets 1 minute 50 seconds, the title slide 40 seconds. A timed rehearsal confirms the count.

What should a research presentation include?

A clear research question, enough method to judge the answer, visible evidence and an honest boundary around the conclusion. In practice that means a title slide that previews the answer, the gap, the design condition behind the result, the main figure, the uncertainty, limitations tied to their consequences, the next test, and where to find the full paper and data.

How much methodology should a conference talk include?

Enough for a specialist to audit the path from measurement to claim: how cases were selected, how development and evaluation were separated, how the endpoint was built, what was compared and which assumptions matter. Apply one test to every cut: could the omitted detail change a reasonable listener's interpretation of the result? If yes, keep its implication in the talk.

How do you prepare for questions after a research talk?

List the likely questions and prepare each in three parts: a short answer, the record that supports it and what remains open. Cover sample size, independence, what the uncertainty estimate includes, generalization and practical significance. Keep backup slides titled by the question they answer, and say plainly when a condition has not been tested.

Can AI turn a research paper into a presentation?

AI can draft the slides once the scientific decisions are made. Give the tool an approved brief with the timing, audience, central claim, retained method, approved figures, data table and overclaims to avoid, rather than only the PDF. Then review the draft against the paper: every number, unit, interval, diagram and claim-bearing title.

Sources

All sources were checked on 7 October 2026. How this guide was made: the method, templates, timing ranges and checklists are editorial practitioner guidance, not a published standard; the sensor study, its data, slide titles and answers are fictional teaching material with real, reproducible arithmetic; conference timings and statistical guidance come from the sources below.

• [1] On-site paper talks of 10 minutes plus 2 minutes of questions at CHI 2026 — ACM CHI 2026, Paper presentation instructions (event 13–17 April 2026).
• [2] Full papers and selected short papers: 12 minutes plus 3 minutes of questions at ETRA 2026 — ACM ETRA 2026, Instructions for authors.
• [3] Redesigning paper figures for slides, replotting selected data and identifying adaptations — Matthew Goss, MIT CEE Communication Lab, Redesigning existing figures for slides (15 January 2024).
• [4] Analyzing naturally paired measurements through their differences — NIST/SEMATECH e-Handbook of Statistical Methods, Analysis of paired observations.
• [5] The t-based confidence interval for a mean and its repeated-sampling interpretation — NIST/SEMATECH e-Handbook of Statistical Methods, Confidence limits for the mean.
• [6] Recording a practice talk and leaving time for audience questions — IEEE Author Center, Present your paper.
• [7] Short presentations need deliberately reduced scope and practice — Christopher J. Lortie, Ten simple rules for short and swift presentations, PLOS Computational Biology (30 March 2017).

Build your next conference talk in NOXI, free: attach your approved brief, figures and data table, match the conference or lab template, and export a PDF and an editable PowerPoint for the session.