The fastest way to revise and publish research.

Enjamb brings your team, comments, and manuscript into one intelligent platform with AI-native workflow automation.

Join waitlist

See features

Trusted by researchers at leading institutions

Built by researchers, for researchers

The Word experience, built for research

Migrate instantly. Import your manuscript and start collaborating in seconds. Our editor feels just like Word and, most importantly, keeps 100% of your .docx formatting perfectly intact.

Anaphora—our state-of-the-art model for academic writing.

Ask Anaphora…

Sure! I found 3 relevant papers in PubMed and 2 in ResearchGate…

Help me address comment #4

Ask Anaphora…

Sure! I found 3 relevant papers in PubMed and 2 in ResearchGate…

Help me address comment #4

Tap into real-time intelligence across 100M+ sources to resolve comments, manage tasks, and find citations.

Intelligent Merge

Collaborate without fear. Anaphora understands what you're writing, not just where. It intelligently combines all co-author edits automatically, eliminating conflicts and broken documents.

Journal Templates

Submit to any journal with ease. Use one of our templates to auto-format your manuscript to fit journal guidelines.

Built by researchers, for researchers

The Word experience, built for research

Migrate instantly. Import your manuscript and start collaborating in seconds. Our editor feels just like Word and, most importantly, keeps 100% of your .docx formatting perfectly intact.

Anaphora—our state-of-the-art model for academic writing.

Ask Anaphora…

Sure! I found 3 relevant papers in PubMed and 2 in ResearchGate…

Help me address comment #4

Tap into real-time intelligence across 100M+ sources to resolve comments, manage tasks, and find citations.

Intelligent Merge

Collaborate without fear. Anaphora understands what you're writing, not just where. It intelligently combines all co-author edits automatically, eliminating conflicts and broken documents.

Journal Templates

Submit to any journal with ease. Use one of our templates to auto-format your manuscript to fit journal guidelines.

Ditch the clutter

The entire research process in one window

Write your manuscript, manage tasks, view comments, and write rebuttals - all in one place.

Comments you upload are automatically converted into actionable tasks and assigned to collaborators based on their skills & role

Confirm the formula for the position-wise feed-forward network

Expand the section on positional encodings to provide more intuition behind the choice of sine and cosine functions

Add a brief note on why d_k and d_v were set to 642

Verify the O(1) sequential operations claim for self-attention in Table 11

Jon Porchet

Confirm the formula for the position-wise feed-forward network

Expand the section on positional encodings to provide more intuition behind the choice of sine and cosine functions

Add a brief note on why d_k and d_v were set to 642

Verify the O(1) sequential operations claim for self-attention in Table 11

Jon Porchet

Verify the 3.5-day training time for the "big" model.

Confirm the P_drop rate of 0.1 for the base model.

Explain intuition behind the learning rate formula & clarify the purpose of warmup_steps before the inverse square root decay begins.

Check the WMT 2014 EN-DE training set size (4.5M pairs).

Saimon Kurō

Verify the 3.5-day training time for the "big" model.

Confirm the P_drop rate of 0.1 for the base model.

Explain intuition behind the learning rate formula & clarify the purpose of warmup_steps before the inverse square root decay begins.

Check the WMT 2014 EN-DE training set size (4.5M pairs).

Saimon Kurō

Clarify the purpose of the decoder's self-attention maskin and explicitly state that the mask prevents positions from attending to subsequent positions.

Note the use of byte-pair encoding for the EN-DE task.

Confirm the parsing results for the WSJ-only 4-layer model.

Check the beam size and length penalty (alpha=0.6) used for translation.

Parbir Pattabiram

Clarify the purpose of the decoder's self-attention maskin and explicitly state that the mask prevents positions from attending to subsequent positions.

Note the use of byte-pair encoding for the EN-DE task.

Confirm the parsing results for the WSJ-only 4-layer model.

Check the beam size and length penalty (alpha=0.6) used for translation.

Parbir Pattabiram

in seconds.

To a clean task-oriented workspace

Go from unorganized comments

Comments you upload are automatically converted into actionable tasks and assigned to collaborators based on their skills & role

Confirm the formula for the position-wise feed-forward network

Expand the section on positional encodings to provide more intuition behind the choice of sine and cosine functions

Add a brief note on why d_k and d_v were set to 642

Verify the O(1) sequential operations claim for self-attention in Table 11

Jon Porchet

Verify the 3.5-day training time for the "big" model.

Confirm the P_drop rate of 0.1 for the base model.

Explain intuition behind the learning rate formula & clarify the purpose of warmup_steps before the inverse square root decay begins.

Check the WMT 2014 EN-DE training set size (4.5M pairs).

Saimon Kurō

Clarify the purpose of the decoder's self-attention maskin and explicitly state that the mask prevents positions from attending to subsequent positions.

Note the use of byte-pair encoding for the EN-DE task.

Confirm the parsing results for the WSJ-only 4-layer model.

Check the beam size and length penalty (alpha=0.6) used for translation.