Paper writing

The Paper Writer is an online, multi-file LaTeX workspace: conference templates, a CodeMirror editor with real-time collaboration, server-side compilation with a live PDF preview, and an AI drafting agent that writes the paper section by section — under hard guardrails that keep every number and every citation tied to something real.

The page lives in the topic workspace under Paper Writer.

How it works

Manuscripts and templates

A manuscript is a self-contained LaTeX project inside a topic: a file tree of .tex sources, references.bib, style files, and uploaded assets. When you create one you pick a template:

• Built-in: NeurIPS 2026, ICLR 2026, and ACL, shipped with the platform. Their style files (.sty/.cls/.bst) are read-only so you cannot accidentally break the venue format.
• Official: additional venue packs downloadable on demand (the picker shows "Official · not downloaded" until fetched).
• Custom: upload any template as a zip ("Upload zip"); the main .tex file is detected automatically and binary assets are stored as read-only files.

Each manuscript has a main file (the compile entry point) and a compiler — tectonic by default (always available in the Docker image), with pdflatex / xelatex / lualatex offered when TeX Live's latexmk is installed on the server. Both are switchable in the editor top bar, Overleaf-style.

The editor

The editor page is three panes: file tree, CodeMirror editor, and the PDF preview.

• Real-time collaboration. Every text file is a CRDT document synced over WebSocket — multiple people (and the AI) can edit the same file at once, with peer cursors and an online-collaborators list in the top bar. Changes snapshot to the database within ~2 seconds, so compiles and the REST API always see the latest text. If your connection drops, edits are kept locally and merged when the socket reconnects ("Connection lost — changes are kept locally and will sync after reconnecting").
• Compile and preview. Press ⌘S or click Compile. Compilation runs server-side with the selected engine and streams back a versioned PDF plus parsed diagnostics — undefined citations, undefined references, LaTeX errors with file:line, and overfull-box warnings. Clicking a diagnostic jumps to the file and line.
• Files. Create files and folders, rename, upload (binary uploads such as images become read-only assets), and pick any .tex as the main file. figures/ is reserved: experiment figures are injected there at compile time.
• Fact pack drawer. The Fact pack button shows the manuscript's fact sources at a glance: the idea, hypotheses with their verified/falsified status, every metric from every experiment run, the QC-checked figures, and the citable knowledge-base entries with their bibkeys.

The fact pack: where the guardrails come from

When a manuscript is created (and automatically refreshed before every AI drafting run) the platform assembles a fact pack from real data:

Section	Source
idea	The linked idea's title and summary
hypotheses	The linked experiment's hypotheses with status and evidence
metrics	Every metric of every ExperimentRun, per run, with the direction-aware best value
figures	The experiment's VLM-checked figures (exp_fig_N, with captions)
citations	The topic's knowledge base: compiled/included library papers plus shelf papers, each with a stable bibkey

This is the only material the drafting agent is allowed to state facts from — see the guardrails below.

AI drafting

Click AI draft to start a drafting task (kind paper_writing; one at a time per manuscript):

1. Initialize structure (step 1 in the dialog, and automatic if you skip it): the document body of your main file is replaced with a section skeleton (% POLARIS_SECTION: ... markers) in a new draft.tex, which becomes the compile main file. Your original file is kept untouched.
2. Pick sections — all of them, or a subset — and add optional notes for the AI.
3. The agent drafts section by section in a fixed order: Introduction → Method → Experimental Setup → Results → Conclusion → Abstract, then compiles, then writes Related Work last, then runs the final compile. The run only completes if the final compile succeeds.

While it writes you can watch it live in the editor: the text streams into the section as the AI "types", with a status cursor showing the phase (typing / revising / compiling). Collaborator edits elsewhere in the file merge normally through the CRDT.

Every section passes three hard static checks before it is accepted:

• \cite{key} may only use bibkeys present in the fact pack's citations — a citation can never be invented;
• \includegraphics may only reference fact-pack figure ids (as figures/<fig_id>.pdf);
• every percentage and decimal in the text must match a real ExperimentRun metric value (±0.01), with small integers, years, and section/table numbers exempted.

A violating draft is rewritten (up to twice, with the violations quoted back). If it still fails, the section is written anyway but prefixed with a % TODO comment listing the problems, and a "needs review" activity is recorded — the run degrades loudly instead of silently shipping an unverifiable claim. Clean sections get one self-reflection polish pass (which must itself re-pass the checks, otherwise the original stands).

Related Work is special: the candidate pool is the fact-pack citations plus the top-10 Semantic Scholar title matches for your paper, and the agent may cite only from that pool. Any S2 candidate it actually cites is added to the fact pack (and references.bib) as a @misc entry, so the compile never hits an undefined citation.

Bibliography, one click

Refresh references regenerates references.bib from the current fact pack — stable citation keys of the form {firstAuthor}{year}{firstWord} — and fixes the \bibliography wiring in the main file if it is missing or points elsewhere. Run it after the knowledge base or the experiment changed; drafting runs do it automatically.

Inline AI assist

For hand-editing, select text in the editor and open the assist panel:

• Polish — improve the selected passage;
• Rewrite — transform it following your instruction ("more concise", "active voice", …);
• Continue — extend from the cursor.

The result streams in beside the editor and is only applied when you accept it. The same fact-pack constraints are given to the model, and the applied text is statically checked afterwards — issues are shown as warnings rather than blocking, since you are in the loop.

arXiv export and submission

• Export arXiv builds a clean, submission-ready tar.gz: sources, references.bib, figures, and a freshly generated .bbl (the server recompiles once with intermediates kept so the .bbl matches the current sources), with build artifacts stripped. Any caveat — e.g. no compiler available to produce the .bbl — is reported as an export note.
• Submit requests submission approval. It requires the latest compile to be ok and the manuscript to have passed paper review; it then opens a paper_submission approval gate. Approval marks the manuscript submitted; rejection returns it to compiled.

Step-by-step usage

1. Open Paper Writer, pick your topic, click New manuscript, choose a title and template.
2. Optionally link the idea and experiment when creating it — that is what populates the fact pack with hypotheses, metrics, and figures.
3. Open the manuscript, click AI draft, initialize the structure, select sections, start.
4. Watch the draft stream in; fix any % TODO(AI …) blocks it left for you.
5. Edit by hand with collaborators; use select-text assist for polish; ⌘S to compile as you go.
6. Refresh references if the library changed; check the Fact pack drawer when in doubt about what the AI is allowed to claim.
7. Run paper review, fix what it finds, then Submit (or Export arXiv).

Key settings

• Main file and compiler — top bar dropdowns; per manuscript.
• Collaborators — manage who can edit and get a share link from the Collaborators dialog.
• Templates — admins and users can upload zip templates; built-in style files stay read-only.
• Pin / trash — manuscripts can be pinned to the top of the list; deletion is a two-step trash + permanent purge, restorable in between.

Tips and limits

• Compilation has a hard timeout; a paper that compiles for minutes usually has a runaway package or figure problem — check the diagnostics panel and trim.
• The drafting agent writes into POLARIS_SECTION markers. If you delete the markers, section drafting falls back to appending at the end of the document — keep them where you want the AI to write.
• Numbers you type by hand are not blocked by the writer — but they will be caught by paper review's fact check if they do not match the experiment record.
• Version history: every AI edit and structural change snapshots the file; use the history dialog on a file to inspect and restore earlier versions.
• One drafting task per manuscript at a time; the top bar shows "AI writing — view task progress" linking to the task console while it runs.