--- name: neurips26-talk description: Plans and builds slides, a timed script and likely questions for a NeurIPS 2026 oral talk from the user's paper. Use when the user is preparing an oral or spotlight talk, slides, or a SlidesLive recording for NeurIPS 2026. license: MIT --- # NeurIPS 2026 oral talk This skill is unofficial. It comes from the Quantitative Research Society (QRS) attendee guide at https://qrsntu.org/service/neurips26/oral/ and is not from NeurIPS. The facts below were checked against the official pages on 5 November 2026. Rules can change, so tell the user to check the official pages before uploading slides. ## Rules (each with its source) - Screens are 16:9 with HDMI input. Bring your own adapter and a laser pointer; NeurIPS does not provide one (https://neurips.cc/FAQ/AudioVisualProjectorConnection). - Do not depend on venue internet. Keep every slide and video on the laptop (https://wiki.eventhosts.cc/en/reference/physical-presenter-instructions). - The first slide names the paper and the authors, or has a QR code to the paper page. SlidesLive matches the recording to the paper from the opening slide (same page). - Upload one PDF of at most 10 MB at https://neurips.cc/SlidesUpload, and bring the same file on the laptop (same page). The 10 MB limit, 16:9 and the title-slide or QR-code advice are general guidance from the conference platform wiki, not year-specific, and unconfirmed for 2026. - Arrive 15 minutes early; the centre of the front row is kept for speakers (same page). - Talk length is not published. In Atlanta four papers share each 60-minute session. Ask the user to check their session on https://neurips.cc/MyStuff, and plan about one slide a minute until they know. - Oral papers also present a poster in their poster session (https://neurips.cc/Conferences/2026/PosterInstructions). If no one gives the talk, NeurIPS may downgrade the paper to a poster or revoke acceptance (https://neurips.cc/Conferences/2026/MainTrackHandbook). - The SlidesLive video is strongly recommended, not required, by 13 November, through Presentation History on MyStuff. ## Accessibility (Main Track Handbook) Source: https://neurips.cc/Conferences/2026/MainTrackHandbook - Minimal text, large fonts, bold for emphasis, no text effects such as shadows. - High contrast; dark text on a cream background works best. - Colour-blind safe palettes; never use colour alone to carry a message. - No flashing text or graphics. - Do not assume everyone can see the slides: say aloud everything important, and describe each figure. - Use examples a diverse, multicultural audience understands; nothing offensive. ## Templates No talk template is on the public NeurIPS pages as of 5 November 2026. QRS templates (unofficial, MIT licence, 16:9), all under https://qrsntu.org/service/neurips26/assets/talk/: - beamerthemeneurips26.sty (theme) and talk.tex (nine-slide example) - neurips26-talk-latex.zip (both plus README.txt) - neurips26-talk.pptx (the same nine slides) - example-talk.pdf (the beamer example as compiled) Commands in talk.tex: `\title`, `\author`, `\institute`, `\paperurl` (the QR code), `\paperid`, `\session`, `\contact`, `\placeholder{height}{label}`, `\takeaway{sentence}`, `\closingslide{sentence}`. Compile with pdflatex, twice. ## Workflow 1. Get the inputs: the paper, the talk length in minutes (or "unknown"), the session line (for example "Atlanta Orals 1A, Wed 9 Dec, 15:30, Hall C3"), and the paper page URL from MyStuff. Write TODO for anything missing. 2. Read the whole paper. Propose an outline at about one slide a minute: title, the problem, why it is hard, our idea, method, main result, what else we found, takeaways, closing. Give each slide one message as its title and a time in seconds. Show it to the user and wait for their OK. 3. Build the slides (output options below). One message per slide, at most about six short bullets, and prefer one figure and one sentence. 4. Use the paper's own figures. For each, write the sentence the speaker will say to describe it: the axes first, then what to look at. 5. Never invent numbers, results or citations. Every number must appear in the paper. Where something is missing, write TODO. 6. Write a speaker script, slide by slide, with a time per slide, that fits the talk length with a minute to spare. 7. Check: 16:9 pages, no overfull boxes, fonts readable from the back of a hall, PDF under 10 MB, first slide has the title and authors or the QR code. 8. Prepare three likely questions with short answers, taken from the paper's limitations and experiments. Other advice, with sources: start around 22 pt sans-serif, about six bullets at most, no all caps, add alt text, use a green rather than red laser pointer (RECSYS accessible presentations guide, linked from the Handbook); put detail in a handout rather than on slides (W3C accessible presentations guide, linked from the Handbook). ## Output options **Beamer .tex + PDF (default).** Edit talk.tex with beamerthemeneurips26.sty and compile. The PDF is the file to upload. **PowerPoint .pptx.** Start from neurips26-talk.pptx. The slide size is 33.867 x 19.05 cm (16:9); do not change it. Export to PDF for the upload. **HTML deck or Claude Artifact.** Write one self-contained HTML file with one `section` per slide. Use print CSS so each slide is one page at 16:9: ```css @page { size: 338.67mm 190.5mm; margin: 0 } section.slide { width: 338.67mm; height: 190.5mm; page-break-after: always } ``` Use large sans-serif type, dark text on cream, embedded images, no external requests, and no animation that flashes. Tell the user to print to PDF at 100% scale, check the page size and file size, and upload that PDF. ## What to hand back The source file, the PDF, the speaker script with timings, the three questions with answers, and a list of every TODO. Remind the user to bring the laptop, an HDMI adapter and a laser pointer, to arrive 15 minutes early, to upload the PDF at https://neurips.cc/SlidesUpload, and to check https://neurips.cc/Conferences/2026/MainTrackHandbook again before uploading.