The question of how to write a paper that resonates with a NeurIPS, ICLR, or ICML audience is one that every researcher in machine learning wrestles with. The truth is, too many submissions read like a technical manual, a dry recitation of methods and results that fails to hold the reader's attention. The community has recently grappled with related pressures, from the sting of Why Top-Tier ML PhDs Face Rejection from Leading AI Internships to the frustrating silence of a Desk Rejection in Two Days: When Rigorous Work Meets Silence. Amidst this environment, Michael J. Black's argument that we should write in a story-oriented rather than results-oriented style is not just a stylistic preference, it is the practical path forward for getting your work read, understood, and accepted.
We agree with Black's core premise, but let's push it further. Story-oriented writing is not about embellishing your findings or fabricating drama. It is about constructing a clear narrative arc that answers three questions: Why should I care? What did you do? And what does this mean for the field? A paper that leads with its results forces the reader to retroactively piece together the motivation and the experimental logic. A story-oriented paper guides them through the problem space, showing the gaps in existing work, the insight that led to your approach, and the concrete evidence that validates it. This structure respects the reviewer's limited time and cognitive load. It does the hard work of making your contribution feel inevitable and important, rather than leaving the reader to guess why it matters.
This approach also inoculates your work against a specific danger we've seen surface this year in reviews: the rise of automated screening tools. As noted in coverage of Navigating AI-Slop Detection in ICLR's First Automated Review Cycle, detection systems are already filtering for generic, template-like language. A story-oriented paper, built on a unique narrative that reflects the specifics of your research process, is inherently harder to flag as formulaic. It demonstrates that a human mind crafted the argument, which is precisely the signal that these systems are trained to value over surface-level compliance with formatting.
The concrete takeaway is this: Before you write a single experimental result, draft your abstract as a four-sentence story. Start with the status quo. Describe its limit. Introduce your key insight that overcomes that limit. End with the single most convincing piece of evidence that your insight works. If you cannot write that story, you do not yet understand your own contribution deeply enough to submit. The reviewers at these conferences are not looking for exhaustive catalogs of benchmarks; they are looking for a reason to believe in your idea. Give them that, and your rigor will speak for itself, rather than being buried under it.