LLMs, Thin Slices, and the Secret to a Great Scientific Presentation

Published on May 5, 2025 4 min read

Delivering a compelling presentation is a key skill for scientists—and the opening minutes often matter most. New research shows that AI can evaluate just the beginning of your talk and still provide meaningful, targeted feedback. It’s a fast, low-effort way to sharpen your delivery where it really counts.

If you’ve ever given a scientific presentation and noticed your audience mentally checking out before you hit slide three—congrats, you’ve experienced the brutal power of the "thin slice". That’s the term psychologists use for the quick judgments people make based on just a few seconds of observation. Turns out, those judgements are often surprisingly accurate, just maybe not in your favor.

When your groundbreaking research is so fascinating that your audience decides it's nap time
When your groundbreaking research is so fascinating that your audience decides it's nap time. Created by ChatGPT and inspired by PHD Comics

For scientists of all levels, those first impressions matter a lot. Whether you’re defending your thesis, giving a job talk to potential employers, or pitching a new project to senior leaders, those first moments on stage are doing a lot of heavy lifting. Like it or not, people often decide whether you’re credible, confident, and competent before your science gets a chance to shine.

Now here’s where things get interesting: A recent study (Schmälzle et al. “The Art of Audience Engagement: LLM-Based Thin-Slicing of Scientific Talks.” arXiv, April 15, 2025) found that large language models (LLMs)—yes, the same tech that you might have used to polish your slides—can predict human ratings of presentation quality just from thin slices. Even better, this works reliably across different LLM models and prompts.

Top: Illustration of how presentation transcripts are divided into thin slices—1%, 5%, 10%, and so on—for evaluation. Bottom: Even with just 10% of the talk, LLM-based evaluations closely match the scores for the full presentation. (Source: Schmälzle et al. 2025)
Top: Illustration of how presentation transcripts are divided into thin slices—1%, 5%, 10%, and so on—for evaluation. Bottom: Even with just 10% of the talk, LLM-based evaluations closely match the scores for the full presentation. (Source: Schmälzle et al. 2025)

In other words, AI doesn’t just help you create your presentation. It can now help you test drive your delivery. Before you face a real audience, you could run a snippet of your talk through an LLM—a fast, low-effort way to gauge your audience’s first impression. Think of it as rehearsal with a brutally honest (but never judgmental!) AI colleague.

You will receive a transcript of a science/research presentation. Rate the overall rhetorical quality on a scale from 1 (worst) to 10 (best), considering clarity, engagement, structure, and delivery. Return only the single rating number as a plain integer, with no other text or characters. Here is the speech text: <presentation snippet>
6 Suggestions for Improvement: Clarity: The speech is clear and informative, but it could benefit from more specific examples or data to support the claims made. Engagement: The tone is factual but somewhat dry. Adding a compelling anecdote, visual imagery, or rhetorical question could enhance audience engagement. Structure: The structure is logical, with a clear introduction and overview of the topic. However, transitions between points could be smoother. Delivery: The language suggests a formal and competent delivery. Using more dynamic language or emphasis might improve retention and interest.

Of course, AI can’t replace storytelling, presence, or human connection—these are skills you build through practice, not shortcuts. So treat AI as an assistant: something to help you refine, reflect, and improve. The goal isn’t to sound like a machine—it’s to become a better version of yourself.

    Scientific PresentationGenerative AILanguage Models