How to explain a research paper in a video

Research papers are written for reviewers. The abstract is dense, the introduction argues with prior work, and the real contribution often sits in one figure on page 5. A good paper summary video does the reader's work for them: it pulls out what the paper did and why it matters, in a couple of minutes, for someone who will never read section 4.

Here's a structure that works for almost any paper, the manual ways to produce the video, and an example made from the GPT-2 paper.

The four-part structure

Most papers, whatever the field, answer four questions. A summary video should answer them in this order.

Problem. What wasn't working, or wasn't known, before this paper? One or two sentences, in plain words. If you need jargon to state the problem, define it here.
Method. What did the authors actually do? Name the idea that's new. Skip the parts that are standard in the field.
Result. What did they find? Pick the one or two results that support the main claim. A single number with context ("beat the previous best on 7 of 8 benchmarks") beats a table read aloud.
Limitation. What doesn't it show? Where does the method break, what's left untested, what do the authors themselves flag? This is the part summary videos most often skip, and the part that makes yours trustworthy.

Applied to the GPT-2 paper ("Language Models are Unsupervised Multitask Learners", Radford et al., 2019), it looks like this. Problem: language systems were trained separately for each task on labelled data, and generalised poorly. Method: train one large language model (1.5 billion parameters) to predict the next word on a big scrape of web pages, then test it on tasks it was never trained for. Result: with no task-specific training, it set new state-of-the-art results on 7 of 8 language modelling benchmarks and showed some ability at reading comprehension, translation and summarisation. Limitation: on most of those other tasks it was still far from usable, and the authors note the model still underfits its training data.

That's about 100 words, which is most of a one-minute script.

A worked example

Here's a Rekka video made from the GPT-2 paper:

Option 1: Script it and narrate over figures

The manual way, and still the right one for a conference talk or anything with your name on it.

Read the abstract, the figures and their captions, and the conclusion first. Then read the method section closely. Most papers can be understood well enough for a summary this way.
Write the four parts as spoken sentences. At about 150 words per minute, a three-minute video is roughly 450 words. Hold yourself to that.
Plan one scene per idea. The paper's own figures are usually the best visuals for the result. For the method, a simple diagram you draw yourself often explains better than the paper's architecture figure, which was made for experts.
Record narration over slides or a screen recording of the figures, then export. PowerPoint and Keynote can both record narration and export to video.

The good: you control every claim, and the process makes you understand the paper properly. The bad: it's hours of work per paper, and the hardest part (turning the method into something a non-specialist follows) is where most time goes.

Option 2: Talk through the paper on screen

Some of the best paper explainers online are just a person scrolling through the PDF, highlighting, and talking. It's quick to make and honest about the source. It also tends to run long, because without a script you explain everything in the order the paper does. If you go this route, write the four-part outline first and stick to it.

Option 3: Generate it from the PDF

A generator can read the paper and produce the summary video, leaving you to check it. For papers, checking is the job: read the script against the abstract and results section, and make sure the numbers and the limitation are right. A summary that overstates a result is worse than no summary.

How Rekka does it

Attach the paper in the chat at rekkaai.com/app and send it. A PDF is the usual format, but Word files, and photos of printed pages, work too. With no message, the video covers the whole paper. You can direct it in the same message:

"Focus on the method in section 3."
"Explain it for a first-year undergraduate."
"Problem, method, result, limitation, in that order."
A language for the narration.

Rekka reads what's on the pages, including the equations, tables and figures, and writes the script from them, so the video uses the paper's own terms and numbers. It then draws the explanation scene by scene in a whiteboard, chalkboard or crayon style while an AI voice narrates. The mp4 shows up in the chat; download it or turn on a public share link. Landscape, square and portrait are all options.

Before spending anything, you can ask questions about the paper in the chat: "what's the main contribution?", "what are the limitations the authors mention?", "explain equation 2". That costs no credits, and it's a good way to check the reading before you make the video.

Practical limits:

Up to 30 pages per message. Most papers fit; for a long one with appendices, send the main body.
One credit makes one video of about 100 seconds, and new accounts get one free credit. On the free plan, the video is a single chapter of about 100 seconds summarising the paper. Paid plans can make longer, multi-chapter videos, and you see the length and credit cost and confirm before it's made.

The research paper to video page has more, and PDF to video covers PDFs in general.

Mistakes that make paper videos worse

Reading the abstract aloud. It was written to be skimmed by experts, and spoken it's a wall of terms.
Explaining the related work. Viewers want to know what this paper did. One sentence on what came before is enough.
Showing every result. Pick the one that supports the main claim.
Skipping the limitation. Viewers can't tell an incremental result from a breakthrough unless you say what the paper doesn't show.
Getting the numbers wrong. Whoever or whatever wrote the script, check every figure against the paper before you share it.