Skip to content
Open the library
Paper2Lab

The lead

Read any research paper at your level.

Paper2Lab explains every section three ways, from new-to-this to expert, and pins each explanation to the sentence in the paper it came from.

Open the libraryRead a sample paper

Plate I. The reader, workingdrag the level

Vaswani et al., Attention Is All You Need§ 3.2

Key terms, briefly explained

Paper2Lab explains

The paper stays on the page. The help goes in the margin.

Paper2Lab never paraphrases a paper out of sight. Here is part of section 3 of Attention Is All You Need, with the notes the reader adds beside it. Hover the equation, follow a quote to its page, answer the question.

Excerpts quoted word for word. Text between them is left out […].

Every claim shows its quote.

Any AI can sound sure of itself. Paper2Lab has to quote the paper for every explanation, and a short Python program, not the AI, looks for each quote in the PDF's own text, word for word. Figure captions are checked the same way.

Beside every quote, one of two marks

Found word for word in the PDF. The quote links to its page, so you can read it in context.

Couldn't find this exact wording in the PDF. A quote that fails the check stays on the page, marked, never hidden.

How accuracy is measured

A small shelf, read closely.

The demo library holds two landmark machine-learning papers, explained in advance so every feature works without an AI key: three reading levels, equations in words, figures, questions with answers, and a map of what to know first.
  • arXiv:____.____

    Your paper, next.

    Paste an arXiv link or drop in a PDF. It's read once, explained section by section, and kept in this browser only. No account.

Equations you can push on.

Some ideas only click when you poke them. Key equations in the samples come with a small demo, right under the formula in the reader.

Turn the key size up and switch the scaling off: the softmax piles all its weight on one word, which is exactly the problem the paper's √dₖ fixes.

See it in the reader
Play with itWhy divide by the square root of the key size?
128

How much “animal” attends to each word

  • The13.1%
  • animal30.2%
  • didn't47.1%
  • cross3.7%
  • street5.9%

Scores q·k/√dk: -0.1, 0.8, 1.2, -1.4, -0.9 · softmax slope at the top word 0.249 (max 0.250)

With scaling, attention is spread out (the top word gets 47.1%), so the gradients are healthy and it can still learn.

Queries and keys are random numbers with mean 0 and variance 1, as in the paper's footnote, so a dot product of dk of them has variance dk.

Notes to the reader

Your papers stay yours

There are no accounts. Papers you add belong to your browser, through a random key in a cookie that scripts can't read. Nobody else can open, list or delete them.

About this demo

The live demo runs without an AI key, so it shows two papers explained in advance. Every feature works on them, including prepared answers to their suggested questions. Adding new papers is paused.

Made in the open

Explanations are written by AI and can be wrong; that's why each one shows its quote. The code is open source, built by Alshifa.

Pick a paper. Pick your level.

Open the libraryHow accurate is it?