6. Writing & Publishing
Good papers take a lot of iteration. Getting something out is real work, so we want to start early, share rough versions quickly, and iterate many times.
Story and figures first #
I really believe in trying to nail down the story and your contributions first. Sometimes that’s hard during open-ended discovery when you don’t know exactly where a project is going yet, and that’s normal. Still, a paper is built out of 4, 6, or 8 figures (see Aesthetics for how we think about figure arcs), and nailing down what those figures are is what actually tells the story.
There’s no single formula, and what works for you may change over time, though here’s a sequence I recommend:
- Mock up the figures first. I’m a big fan of doodling on paper and taking screenshots. You can also use GenAI image tools to mock up a rough sense of how a visual layout might look (one of the few areas where I think GenAI is genuinely useful for images).
- Build a skeleton of the story beats: Introduction, Methods, Experiments and Results, Discussion, and the main points and contributions we want to cover in each.
- Write a 2 to 3 sentence pitch for your idea.
- Write bullet points inside each section for what you expect to cover.
- Fill in text, and repeat.
The most common mistake I see on a first paper draft is waiting until everything feels done and perfect before showing anyone. Show drafts really early. It’s okay if it’s rough: iterating together is how it gets good.
For tools, I recommend Google Docs for drafting (fast, simple, collaborative) with Paperpile for references: copy a DOI and let the software figure out the rest. When we’re close to submission, we export with BibTeX to LaTeX on Overleaf.
You can also write with other people. Pair writing works much like pair programming, since talking through your thoughts with someone else puts you in a different mode. Communal writing sessions add a bit of social accountability: set a concrete goal (“finish half a page of the introduction”), write for a fixed block of time with breaks, share progress at the end, and talk through whatever is making the section hard.
Resources:
- Step-by-Step Guide to Undergraduate Writing
- How to write a first-class paper (Nature)
- Preparing manuscripts: scientific writing for publication
- Rubber duck debugging and pair programming
- Strengthening the case for pair programming
- Pair writing
- U of T Writing Centres, a great resource
Feedback, internal reads, and user journeys #
I tend to give writing feedback directly on the document or during our weekly 1:1s. If I’m busy or we need a longer pass, book a dedicated slot on my booking page [TBD] just to go over the draft together.
Ideal lead times once it’s clear you’re aiming for a deadline:
- Workshops and conferences need at least two weeks, with multiple iterations.
- Journal papers need at least a month to polish figures and writing.
Before a paper goes out, everyone should do internal reads with labmates. Presenting your draft story in our weekly group meeting is also one of the best ways to practice and catch holes early.
If the project involves software, run a user journey before release: ask someone unfamiliar with the codebase to spend 30 to 60 minutes running the main workflow you expect most people to use, and have them write down quick impressions in a doc of every spot that felt rough. AI agents make code much more replicable now, though we’re building software for people first, so we still need to feel out how a newcomer actually experiences it.
Authorship #
Authorship is a conversation to have early and often. It can be uncomfortable, so it’s much better to be honest up front and not have regrets later. Having explicit author contribution statements (like CRediT) is great practice.
I encourage multiple first authors. Spread the wealth and be generous, while staying fair: we don’t add people just because they could be there. To be generous, you have to create those opportunities in the first place. Co-first author order can be decided loosely (even if it shouldn’t matter, the first position does). The last author is usually the PI, reflecting funding and supervision. In collaborations, that’s a discussion between the PIs.
Where we publish #
Choosing a venue is contextual and depends on the work and the deadlines. Our archival venues are AI conferences (such as NeurIPS, ICML, and ICLR) and journals at the intersection of chemistry and machine learning (such as Nature Machine Intelligence, Chemical Science, Digital Discovery, and Journal of Chemical Information and Modeling).
Because our work often gets read by both ML reviewers and bench chemists, write as simply and plainly as you can. Most people enjoy clear, simple writing: even when the ideas are technical, distill why we’re tackling the problem into straightforward language, and make sure anyone looking at a figure can tell right away what claim you’re making.
ML workshops (at NeurIPS, ICLR, or ICML) are great milestones and stepping stones toward a full paper. Just keep in mind that workshops are usually non-archival unless they announce otherwise.
Preprints and scooping #
I recommend always preprinting. It’s better for visibility, citations, and open access, and it’s a good milestone. Post when you’re close to submission and confident. The only risk I see is a false sense of being done.
Scooping happens because good ideas occur to several people at once, and independent work also serves as replication. A nine- or ten-page paper only captures one angle of a problem anyway, so there’s usually room for multiple perspectives on the same idea.
Reviews, rebuttals, and rejections #
For peer review and rebuttals, block that time off on your calendar ahead of time (for AI conferences, the rebuttal window is usually about a week). Guard your time during rebuttals: you can’t overdo everything, and you can’t make everybody happy.
Papers get rejected all the time for different reasons. Sometimes you get a reviewer who doesn’t vibe with your work, and sometimes the work just isn’t mature enough yet. Treat reviews as a chance to get feedback and sharpen the next version.
Code, data, and explainers #
When a paper is preprinted and accepted, make sure the project has a clean, replicable GitHub repository and that the data is easy for others to grab and use.
I also really encourage writing a short blog tutorial or interactive explainer alongside a paper, and sharing it on LinkedIn or other socials. Ask yourself: what is the story you didn’t get to tell because of the academic paper format? AI tools make building interactive pieces easier now, though you still have to think carefully about how to distill the ideas. It’s a great exercise in telling your story in a way that’s more personable and a little bit more free.