Manuscriva
Start free

Your research,
together.

Write with your team, check every citation against its source, and ask a shared AI assistant that only uses the research you choose.

University email verification required for free runs. Phone verification can be completed afterward.

Transformer reading group
Example project
AMKOLB

Project library

Review draft Attention Is All You Need.pdf wmt14-bleu-table2.csv Citation audit

Review draft · 3 editors live

Attention without recurrence

On WMT 2014 English-to-German, the big model reached 28.4 BLEU [1]. It outperformed every previously reported model on both translation tasks.

BloubContradicted · Table 2, p. 8

English-to-French BLEU · bars start at 40

Suggested It set a new state of the art on English-to-German and the best single-model score on English-to-French, where the ConvS2S ensemble still scored higher (41.29 vs 41.0 BLEU).

Write togetherReview citation evidenceInspect your dataShare project context

Check the reference.
Then read the evidence.

A real reference can still be cited for something it never says. Manuscriva checks the record and the claim separately, and shows you the passage or table behind every verdict.

In our demo it caught a review claiming the Transformer beat every earlier model. The paper’s own Table 2 says otherwise for English-to-French.

See the full citation audit
Saved example check
NeurIPS 2017

Attention Is All You Need

Vaswani, Shazeer, Parmar et al. · arXiv 1706.03762

  1. Supported

    The big model reached 28.4 BLEU on WMT 2014 English-to-German, more than 2 BLEU above the best earlier results.

  2. Contradicted

    It outperformed every previously reported model on both translation tasks.

  3. Supported

    Training took 3.5 days on eight P100 GPUs.

  4. Unable to assess

    Convolutional models had already reduced sequential computation before this paper.

Reference records come from Crossref and OpenAlex. A missing record never means a reference is fake.

Inspect your data.

Open your spreadsheet, choose columns and units, and record how the data is prepared. You choose what error bars represent and whether repeated observations should be combined.

Keep the original file alongside your choices so your team can trace every number in the draft back to its source.

Publication figures: coming soon
wmt14-bleu-table2.csv
From Table 2
WMT 2014 BLEU scores
ModelEN-DEEN-FR
MoE26.0340.56
Deep-Att + PosUnk Ensemble—40.40
GNMT + RL Ensemble26.3041.16
ConvS2S Ensemble26.3641.29
Transformer (base)27.3038.10
Transformer (big)28.4041.00

Original file, selected columns and a saved preparation record.

A project your whole team can follow.

One project holds the draft, the paper, the data and the audit. Trying the tabs does not use any of your free runs.

Attention without recurrence

Vaswani et al. replace recurrence and convolution in sequence transduction with attention alone. We summarise the architecture, check the headline translation results against the paper’s own tables, and note what the evidence does not yet show.

Live co-editing · 3 editors
Open the interactive demo

One library.
Every project decision.

Keep papers, datasets, figures, manuscript versions and notes organized by project. Search and filter the files, open their file details, and keep the originals.

Your shared assistant uses the material you choose and recent questions and answers in the shared conversation. Your team can see who asked each question and which passages were included.

Transformer reading group
Review draftManuscript · 3 editors
Attention Is All You Need.pdfPaper · 11 pages · 32 references extracted
wmt14-bleu-table2.csvDataset · 10 rows from Table 2
Citation auditReport · 4 claims checked

Try it free: three runs on us.

  1. Create your account.

    Enter your university email and a valid phone number.

  2. Verify your mailbox.

    Verify your email to unlock three free runs once your university is approved. We review universities that are not yet listed.

  3. Choose what to run.

    Ask the assistant or check your references. You can complete phone verification later. Keep editing and reading your work after your free runs are used.

Room for your next project.

Pick the projects, collaborators and monthly runs your team needs. Not ready to pay? Start with three free runs.

Student

DZD 1,000 / month

  • 3 active projects
  • 2 editors, including the owner
  • 2 additional read-only collaborators
  • 5 citation checks / month
  • 100 assistant replies / month
  • 1 GB private storage

Lab

DZD 20,000 / month

  • 100 active projects
  • 5 editors, including the owner
  • 15 additional read-only collaborators
  • 100 citation checks / month
  • 2,000 assistant replies / month
  • 20 GB private storage

Prices are set separately for Algeria and international customers. Choose a plan, then message us on WhatsApp to arrange payment. Your plan starts once we confirm your payment.

Questions about Manuscriva.

Make room for careful research.

Bring your manuscript and sources into a project your team can share.

Start with 3 free runs

Example content is an independent review of Vaswani et al. (2017). Manuscriva is not affiliated with the paper’s authors or their employers.