papers

read any research paper,
clearly.

every arXiv paper, redrawn for your phone. highlight anything you don't get and an AI explains it.

Coming soon to the App Store

tap a card

not a pdf.
the paper, redrawn.

papers rebuilds every paper as real text on your phone: typeset maths, sharp figures, tables that scroll. tap a citation to see it, jump to any section, pick white, sepia or black.

‹ ☰ PDF Aa ✦

Aa Language

Attention Is All You Need

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, Illia Polosukhin

arXiv:1706.03762 · 12 Jun 2017

3.2.1 Scaled Dot-Product Attention

We call our particular attention "Scaled Dot-Product Attention" [2]. The input consists of queries and keys of dimension d_k, and values of dimension d_v.

Attention (Q, K, V) = softmax (\frac{Q K^{T}}{\sqrt{d_{k}}}) V

(1)

The two most commonly used attention functions are additive attention [2] and dot-product attention, shown in Figure 2.

highlight it.
understand it.

select any sentence and tap explain. or ask for a summary, the key ideas, the limitations. the AI reads the whole paper, and looks things up on the web when it needs to.

3.2.2 Multi-Head Attention

We found it beneficial to linearly project the queries, keys and values h times with different, learned linear projections.

✦

Ask about this paper Gemini · gemini-3.8-flash

SummarizeExplain simplyKey ideasLimitations

Explain: “linearly project the queries, keys and values h times”

Thinking

Instead of one big attention, the model runs h small ones side by side. Each gets its own learned "lens" (a projection) on the same words, so one head can track grammar while another tracks meaning. Their answers are then joined back together.

name a topic.
get the papers.

the learn tab searches arXiv for you and builds a reading track, from the first idea to the latest. every paper is checked against arXiv, so nothing is made up. save it, and read it step by step.

‹Diffusion models

i want to learn about diffusion models

⌕ Looked up “denoising diffusion”

⌕ Looked up “score-based generative models”

Here's a path from the core idea to modern image models.

Diffusion, from scratch

5 papers, from the first idea to today's image models.

1Deep Unsupervised Learning using Nonequilibrium ThermodynamicsWhere the idea starts.

2Denoising Diffusion Probabilistic ModelsThe paper that made it work.

3High-Resolution Image Synthesis with Latent Diffusion ModelsHow Stable Diffusion got fast.

Save to Reading

all of arXiv, in colour.

follow the fields you care about. each one has its own colour, and every paper gets its own art, so no two look the same.

your key stays on your phone.

papers has no account and no servers of its own. you bring the AI.

bring any AI

paste a key from any of these and papers works out which one it is. or point it at any OpenAI-compatible server.

GeminiClaudeOpenAIOpenRouter GroqxAIDeepSeekMistral

kept in the keychain

your key lives in the iOS keychain, the same encrypted vault as your passwords. it's only ever sent to the provider you picked.

nothing collected

no sign-up, no analytics, no tracking. your reading list and progress stay on your phone.

papers is coming to the app store.

free to download. bring your own AI key (Gemini has a free one).