Free arXiv newsletter for researchers

In which the morning's research is cut to size · — · No. 03

New arXiv papers, hand-cut to your interests.

An editorial weekly daily, machine-built and editor-curated by you. Pick your categories; receive a short edition each morning; reply with numbers to get the PDFs.

Pick the arXiv categories you follow. Each morning at 7 a.m. you get a short edition of new papers that match — titles and abstracts translated to Spanish, with the originals one link away.

Start subscription → See today's edition ↓

01 — Today's edition

four sample mornings · real arXiv data
Browse a sample morning
4 sample mornings · real papers, real dates
The Daily Abstract

Hoy en arXiv

cs.LG · cs.AI · stat.ML
Muestra·18 May 2026
Hoy 4 papers nuevos. Responde con los números para recibir los PDFs adjuntos.
Editor's Pick
01

Attention Is All You Need

Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, Polosukhin
cs.LG · keyword match: transformer

The Transformer architecture relies entirely on attention mechanisms, dispensing with recurrence and convolutions. We show superior quality on machine translation while being more parallelizable and requiring significantly less time to train than recurrent or convolutional baselines.

arxiv.org/abs/1706.03762 · responde con 1 para el PDF
También hoy
02

Sparse-Reward RL Without Manual Shaping

Wu, Chen cs.LG · cs.AI

We propose an intrinsic motivation signal that allows agents to learn from sparse rewards without engineered reward shaping, matching or surpassing state-of-the-art on a 26-game Atari benchmark suite.

arxiv.org/abs/2403.21712·responde con 2
03

Retrieval-Augmented Decoders Forget Less Than You Think

Ortega, Lin, Suzuki cs.CL

Long-context retrieval-augmented decoders preserve more fine-grained factual content than equivalently-sized parametric models, even when the retrieved passage is paraphrased rather than verbatim.

arxiv.org/abs/2405.04219·responde con 3
04

Compositional Sample Efficiency in MoE Language Models

Watanabe, Hassan cs.LG

Mixture-of-experts language models trained on a carefully balanced curriculum acquire compositional skills with up to 4.3× fewer tokens than a dense parameter-matched baseline.

arxiv.org/abs/2405.11423·responde con 4
+ 11 papers omitted past your daily cap of 4.
The Daily Abstract
Responde a este correo con los números de los papers que quieres leer.
El listener te enviará los PDFs adjuntos en ~30 segundos.
Suscrito como you@university.edu. Manage subscription · Unsubscribe
Each subscriber's edition is filtered to their own categories — these four are popular presets

01½ — Reply with numbers, get PDFs

our one trick · no competitor has it

Every digest carries a Reply-To address. Type the integers you want, hit send, and the PDFs come back as attachments — usually inside thirty seconds.

02 — Subscribe to your personalized arXiv digest

double opt-in · free · no tracking

Your edition— pick at least one category

No categories selected yet. Try a preset above.
papers per day · the editor's cut
Computer Science · 40
Economics · 3
Electrical Engineering and Systems Science · 4
Mathematics · 32
Astrophysics · 6
Condensed Matter · 9
General Relativity and Quantum Cosmology · 1
High Energy Physics - Experiment · 1
High Energy Physics - Lattice · 1
High Energy Physics - Phenomenology · 1
High Energy Physics - Theory · 1
Mathematical Physics · 1
Nonlinear Sciences · 5
Nuclear Experiment · 1
Nuclear Theory · 1
Physics · 22
Quantum Physics · 1
Quantitative Biology · 10
Quantitative Finance · 9
Statistics · 6
Advanced — keyword filter, daily cap

Comma-separated. Leave empty to include all papers in your categories.

No password·One-click unsubscribe in every email·Source on GitHub

03 — How it works

four steps · zero passwords

  1. You subscribe.

    Pick categories — by preset or à la carte from the 155 official arXiv codes. Optionally narrow by keyword. We send a confirmation email; one click activates.

  2. We ship at 7 a.m.

    One nightly fetch to the arXiv API for the union of every subscriber's filter, then per-subscriber slicing. Titles and abstracts translated to Spanish via MyMemory; originals one link away.

  3. You reply with numbers.

    Type 1 3 5, or "quiero el 2 y el 7", or "the second one". The listener parses the integers, fetches those PDFs from arXiv, and replies with them attached — usually within thirty seconds.

  4. Manage anytime.

    Every digest carries a signed link to edit categories, change the daily cap, or unsubscribe. No login, no password. The link is yours alone.

04 — Frequently asked

four questions, four answers
What does it cost?
Nothing. The whole stack runs on Oracle Cloud's always-free tier (one shared CPU, 1 GB RAM, six Docker containers in ~165 MB idle), plus free quotas of Brevo for outbound mail and Mailsac for reply parsing. If it ever gets popular enough to outgrow that envelope we'll add a Pro tier for power features — but the basic digest stays free, forever.
The reply-to-PDF thing — how does that actually work?
Every digest is sent with Reply-To pointing at a Mailsac inbox we poll every 10 minutes. When a reply arrives, we identify you by your From address, pull up your snapshot of that morning's edition, extract integers from the body (any natural-language wrapper is fine), download those PDFs from arXiv, and reply with them attached. Senders we don't recognise are silently dropped.
Who's behind this?
An indie open-source project. Source on GitHub, MIT licensed, self-hostable end-to-end. Originally built as a university Cloud Computing project; now run as a public service. No company, no investors, no ads, no data resold.
How do I cancel?
Every email has a one-click unsubscribe link in the footer, plus a List-Unsubscribe header so Gmail's native button works too. You can also reply with STOP, or write to the address shown in the colophon.
Colophon Set in Georgia, with system sans for interface chrome and a monospaced stack for code and paper identifiers. Sole accent: amber #D99C5E. Built and served on an always-free VM. The masthead is set 14% letterspaced because newspapers do that and it works.