From the first paper to your next submission

Your AI research companion

Understand papers, keep up with your field, and check your references before submitting.

AI paper summaries, paper alerts and tools built around your scientific workflow.

Every week, free

This week in research

Rising topics

On arXiv, papers submitted the week of 2 November

  1. Sudden rise Faster ways to analyze changing networks ×3.3 compared with usual · 85 papers
    7.7 per 10,000 papers7.5 per 10,000 papers5.3 per 10,000 papers3.8 per 10,000 papers4.1 per 10,000 papers4.2 per 10,000 papers7.5 per 10,000 papers4.3 per 10,000 papers5.6 per 10,000 papers7.6 per 10,000 papers6.1 per 10,000 papers6.3 per 10,000 papers5.3 per 10,000 papers6.7 per 10,000 papers8.6 per 10,000 papers10.4 per 10,000 papers4.0 per 10,000 papers8.8 per 10,000 papers7.5 per 10,000 papers6.9 per 10,000 papers8.5 per 10,000 papers3.4 per 10,000 papers7.9 per 10,000 papers11.5 per 10,000 papers7.9 per 10,000 papers13.6 per 10,000 papers27.4 per 10,000 papers
  2. Comeback Changing AI behavior through its internal weights ×3.8 compared with usual · 19 papers
    6.4 per 10,000 papers12.6 per 10,000 papers4.0 per 10,000 papers5.1 per 10,000 papers5.4 per 10,000 papers2.8 per 10,000 papers3.0 per 10,000 papers4.3 per 10,000 papers7.0 per 10,000 papers0.0 per 10,000 papers0.0 per 10,000 papers4.7 per 10,000 papers4.0 per 10,000 papers10.7 per 10,000 papers5.8 per 10,000 papers5.9 per 10,000 papers5.3 per 10,000 papers0.0 per 10,000 papers3.8 per 10,000 papers5.8 per 10,000 papers5.3 per 10,000 papers10.3 per 10,000 papers6.9 per 10,000 papers3.1 per 10,000 papers6.8 per 10,000 papers0.0 per 10,000 papers20.8 per 10,000 papers
  3. Sudden rise Making AI cheaper with lower-precision numbers ×2.7 compared with usual · 30 papers
    2.6 per 10,000 papers7.5 per 10,000 papers9.3 per 10,000 papers12.7 per 10,000 papers2.7 per 10,000 papers5.6 per 10,000 papers6.0 per 10,000 papers5.7 per 10,000 papers5.6 per 10,000 papers4.6 per 10,000 papers7.7 per 10,000 papers4.7 per 10,000 papers10.6 per 10,000 papers4.0 per 10,000 papers4.3 per 10,000 papers7.4 per 10,000 papers6.6 per 10,000 papers12.6 per 10,000 papers2.5 per 10,000 papers8.1 per 10,000 papers13.8 per 10,000 papers12.8 per 10,000 papers13.9 per 10,000 papers15.7 per 10,000 papers10.2 per 10,000 papers16.7 per 10,000 papers32.9 per 10,000 papers
All the rising topics, field by field →

Weekly letters

One article per field, 16 November 2026

The 11 letters of this week →

Free to start

Research tools ready to use

All research tools
  • Summaries

    Key ideas, results and limits of any paper, in 16 languages.

    No account needed
  • Chat with papers

    Ask a question to one paper or several, answered from the text.

    No account needed
  • Daily paper alert

    New papers in your fields every morning, with short summaries.

    Free with an account
  • Weekly letters

    Each week, the papers that matter in your field, put into perspective.

    No account needed
  • Rising topics

    The topics rising on arXiv, bioRxiv and medRxiv, field by field.

    No account needed
  • Saved searches

    Save a search and get its new papers by email or RSS.

    Free with an account

Your next reading habit

Science worth reading, directly in your inbox

Get new papers with the paper alert, and take a step back with the weekly letters. One field, two complementary formats.

Free with an account

Start with a selection of categories in your field, then refine your paper alert whenever you want.

2. Choose what lands in your inbox

Your choices are kept while you create an account or sign in. Unsubscribe at any time.

What lands in your inbox

Real, recent examples from the field “Artificial intelligence”.

Daily paper alert Artificial intelligence
  1. cs.AITabular Population Priors as Structured Retrieval for EHR Medication Set PredictionAnimesh Agarwal, Meysam Ghaffari, Nina Fatehi, Carlos Morato

    The authors replace searches for similar patient records with population-level statistics about which medication classes tend to occur with particular conditions. On a test set of electronic health records, this change improves predictions without requiring extra language-model calls or a …

  2. cs.AILoRAWeaver: Directional Adapter Routing in LoRA Representation SpaceInish Melroy D Souza, Tom Jackson, Vitor Castro, Rebecca Higginson, Andrea Solt…

    The authors propose a way to choose which task-specific language-model add-on to use for a new request, without training a separate chooser. It performs better than other methods that also require no training across different tasks, and comes close to …

  3. cs.AIDoes the Ranking Hold? A Practitioner-Facing Audit of Hallucination Detectors on Small, Quantized, On-Device …Shubhra Mittal, Gaurav Naresh Mittal

    The authors test several ways to detect incorrect answers from small language models that could run on everyday devices. Detector rankings change between models, and a simple word-based method performs about as well as more complex methods; they also find …

Every weekday, the new papers in your categories, each with a short summary. See a whole paper alert →

Weekly letter Artificial intelligence · 16 November 2026 · 5 min read
Artificial intelligence this week: What impressive scores leave out

New preprints examine the distance between headline performance and useful knowledge, while showing how simulations and physical models can help when data are scarce.

6 papers put into perspective

  1. A virtual cell can look right while missing the response
  2. Simulated outbreaks help most when local records are thin
  3. The hidden language surcharge
  4. Stricter instructions are not necessarily better judgment
Read the whole letter →
Receive them in my field

See what you can do

A useful next step, made concrete

Illustrative previews based on a published paper. Explore a live summary or set up your own paper alert to try the tools.

01 / Understand

Attention Is All You Need

Vaswani et al. · 2017 · arXiv:1706.03762

The question
Can attention alone support a strong sequence-to-sequence model?
The main idea
The Transformer replaces recurrence and convolution with attention. The paper evaluates it on translation tasks.
A point to check
Results on these benchmarks do not establish performance on every language or task.

Editorial reading example based on the abstract. Read the source ↗

Open a ready-to-read AI summary →

02 / Check references

A preprint has a published version

Your reference

Vaswani et al., 2017
arXiv:1706.03762

Publication to review

Advances in Neural Information Processing Systems 30 (2017)

Review the published record and update your bibliography when appropriate.

Illustrative check, verified against the proceedings. Publication record ↗

Explore Bibliography Check → Premium

For the research you do every day

Go further with Premium

All research tools
Premium+

Explore a research question

Ask a question. Get 15 to 30 founding and recent papers, grouped by approach, with a synthesis that cites them.

Literature review

Start from a question
Premium

Watch research close to yours

Describe your project. Get an email when new arXiv, bioRxiv or medRxiv papers overlap with it.

Scoop Watch

Discover Scoop Watch

Check your bibliography, find related work missing from your draft, and monitor research close to your project.

  • Scoop Watch and a paper alert ranked for your research
  • Our most capable model and no daily summary limit
Billing currency

US$9 / month

See Premium tools

Cancel anytime in one click
Payments handled by Stripe

Grounded in scientific literature

Research papers, with their sources

arXiv · Journals via DOI · bioRxiv · medRxiv · HAL · OSF · PubMed

When a license permits it, summaries use the full text. Otherwise they use the abstract and say so. AI output is a starting point: check the original paper for your research decisions.