The choice facing the researcher preparing for CVPR is not just about statistics, it is about what we, as a field, genuinely value. This contributor has solid results, reproducible code ready for release, and a clear understanding of their computational limits. Yet they feel torn between investing scarce cycles into multi-seed reruns or into clearer writing and presentation. That tension reveals a deeper problem: our peer-review culture often demands statistical rigor that is inaccessible to researchers without institutional backing, while simultaneously claiming to reward reproducibility and transparency. We have seen this pattern before, from the Rethinking the compute demands behind LLM post-training research discussion to the NeurIPS workshop notifications delayed leaving submitters waiting for updates saga, where infrastructure limitations and procedural bottlenecks repeatedly penalize those outside well-resourced labs. It is time to say plainly what many whisper: a single-seed result with released, verifiable code is a valid, honest submission.
The practical stakes here are concrete. Our researcher reports that a single set of runs already required enormous effort on slow, unreliable systems. Asking for mean-plus-std across multiple seeds is effectively asking for weeks or months more of work, time that could instead refine the narrative, improve the figures, or strengthen the discussion of limitations. The field has not yet grappled with what this demand means for equity. Researchers at compute-rich institutions can produce multi-seed results almost by reflex; those without that privilege must choose between statistical completeness and actually communicating their contribution. The Explore the future of AI in education as NeurIPS Education Track decisions near conversation reminds us that the community is already rethinking access and inclusion in other dimensions. Why not extend that same thinking to the basic mechanics of experimental reporting?
Our position is direct: the researcher should skip the reruns, invest the saved time in writing and presentation, and submit with confidence. A single-seed result is not a flaw when the code is public, the methodology is clear, and the limitations are acknowledged. Reviewers who penalize such submissions are enforcing a norm that privileges infrastructure over insight. The field should reward the researcher who can explain why their single result matters more than the one who can afford to run the same experiment five times. What we should watch for is whether CVPR and similar conferences formally address this disparity in their review guidelines. Until they do, every researcher in this position faces a quiet test of integrity: do you chase the appearance of rigor, or do you submit the honest work you have? We know which choice serves science better.