The transition from research to communication. You've spent ten weeks exploring interpretability methods and applying them to your concept. Now it's time to communicate your findings in the form of a research paper suitable for submission to a machine learning venue like NeurIPS.
You might imagine your readers as scholars with plenty of time on their hands, carefully poring over your manuscript in quiet contemplation. The reality is more like a large, crowded marketplace. Reviewers are busy. Readers are skimming. Everyone is overwhelmed with papers to read.
Understanding this reality should shape how you write. Your paper competes for attention with hundreds of others. You must make it easy for anyone to quickly tell:
Different readers engage with your paper at different levels of depth:
| Level of Engagement | Number of Readers | What They Need |
|---|---|---|
| Glance at title | Many | Clear, informative title |
| Skim abstract | Moderate | Concise summary of contribution |
| Look at figures and captions | Moderate | Self-contained visual story |
| Read every word | Few | Full technical detail |
The "read every word" readers are your most important ones—they include your reviewers. But you should make the paper work for all levels. The figures and captions alone should tell the main story.
Ted Adelson, a renowned vision scientist, offers a simple formula that captures the essential structure of a good paper:
The Adelson Formula:
This formula works because it answers the reader's natural questions in sequence: What problem? Why is it hard? What's your approach? Why should I believe it works?
Jim Kajiya, former SIGGRAPH papers chair, put it simply:
You must make your paper easy to read. You've got to make it easy for anyone to tell what your paper is about, what problem it solves, why the problem is interesting, what is really new in your paper (and what isn't), why it's so neat. And you must do it up front. In other words, you must write a dynamite introduction.
The introduction is where most papers succeed or fail. A reader who is confused or bored after the introduction will not give your methods section a fair reading.
Opening hook: Why does your concept matter? For interpretability papers, this often connects to AI safety, scientific understanding, or practical applications.
The gap: What don't we currently understand? What has previous work failed to address?
Your contribution: What do you do? State it clearly and specifically.
Preview of findings: What did you discover? Don't make readers wait until page 6 to learn your results.
Paper roadmap: Brief guide to what follows (optional but helpful).
An underutilized technique: explain your main idea with a simple, toy example early in the paper. Before diving into the full complexity of your method, show readers a minimal case that captures the essence of what you're doing.
For interpretability research, this might mean showing your technique on a single neuron, a single prompt, or a simplified model before presenting the full results. Readers who understand the toy example will follow the complex version much more easily.
A typical interpretability paper follows this structure:
| Section | Length | Purpose |
|---|---|---|
| Abstract | 150-250 words | Complete summary: problem, approach, results, implications |
| Introduction | 1-2 pages | Motivation, gap, contribution, preview of findings |
| Related Work | 0.5-1 page | Position relative to prior work; what's new |
| Methods | 1-2 pages | Model, dataset, techniques, evaluation metrics |
| Results | 2-3 pages | Findings with visualizations and quantitative analysis |
| Discussion | 0.5-1 page | Interpretation, limitations, broader implications |
Gone are the days of "We think this is a great idea and we expect it will be very useful. See how it works on this meaningless, contrived problem?" Modern ML venues expect rigorous experimental validation:
End with conclusions about what your work means, what it opens up, or how it changes how we think about the problem. Summarize your key findings and their implications.
Avoid weak "Future Work" sections. A list of things you wanted to do but couldn't get to work reads as: "Here are all the ideas we didn't have time to pursue before the deadline." You get no credit for things you didn't do. If you have genuinely promising directions, mention them briefly in the discussion, but don't end your paper with a to-do list.
Donald Knuth offers the most important principle of good writing:
Perhaps the most important principle of good writing is to keep the reader uppermost in mind: What does the reader know so far? What does the reader expect next and why?
Treat the reader as you would a guest in your house. Anticipate their needs. Would they like some context before this technical section? Perhaps now, after the heavy math, they'd appreciate a concrete example? Guide them through your paper as a gracious host guides a visitor through their home.
Strunk and White's famous advice: Vigorous writing is concise. A sentence should contain no unnecessary words, a paragraph no unnecessary sentences. This requires not that you make all sentences short, but that every word tell.
Wordy vs. Concise:
| Wordy | Concise |
|---|---|
| "the question as to whether" | "whether" |
| "there is no doubt but that" | "doubtless" |
| "this is a method which" | "this method" |
| "in a rapid manner" | "rapidly" |
Cutting wordiness benefits you twice: you have more space to tell your story, and the text is easier for readers to understand.
Many readers will skim your paper by looking only at the figures and captions. Make this path through your paper work:
Knuth's advice: Many readers will skim over formulas on their first reading. Therefore, your sentences should flow smoothly when all but the simplest formulas are replaced by "blah."
David Mermin's "Good Samaritan rule": When referring to an equation, identify it by a phrase as well as a number. Don't write "substituting (2.47) into (3.51)..." when you can write "substituting the loss function (2.47) into the gradient expression (3.51)..."
When discussing related work, be generous. Acknowledge the strengths of prior approaches before explaining how yours differs. Write from a position of security, not competition.
Example of gracious related work discussion:
"A number of papers to be published this year, all developed independently, are closely related to our work. The idea of [approach X] has been proposed by several authors [9, 1, 11] (in particular, see the elegant paper by Hertzmann et al. [11]). The reader is urged to review these works for a more complete picture of the field."
There are perceived pressures to oversell, hide drawbacks, and disparage others' work. Don't succumb. This is in both your long-term and short-term interests.
Reputation matters. "Because the author was [trusted researcher], I knew I could trust the results." Develop a reputation for being clear and reliable. The benefit compounds over your career.
With acceptance rates around 25%, reviewers are looking for reasons to reject. Avoid giving them easy ones:
The Cockroach: You try, but you can't find a way to kill this paper. Nothing too exciting, but well-written, decent reviews, incremental improvement. These often get accepted.
The Puppy with Six Toes: A delightful paper with some easy-to-point-to flaw. The flaw may not be important, but it makes it easy to reject despite the freshness and originality. Many of these get rejected. If you have a rejected "puppy," address the flaws and resubmit—it may become an oral presentation next time.
Interpretability papers have some special considerations:
Reviewers will ask: Why does localizing this concept matter? Be prepared to answer:
Interpretability research is sometimes criticized for "just finding patterns." Your paper needs strong validation:
Your title is the first (and sometimes only) thing people see. Make it count.
Consider the difference:
Was: "Shiftable Multiscale Transforms" (technical but obscure)
Should have been: "What's Wrong with Wavelets?" (poses a question, creates intrigue)
Good titles for interpretability papers often:
Due: Thursday of Week 11
Submit your paper's introduction and methods sections. The introduction should clearly motivate your research question and position it in the literature. The methods section should be complete enough that a reader could reproduce your experiments.
Format: NeurIPS style, approximately 2-3 pages for these sections. (The full paper is 8 pages, with results being the largest section.)