Three papers that changed how I read data this spring

@misc{reading-list-spring,
  author = {Mara Ellison Voss},
  title = {Three papers that changed how I read data this spring},
  howpublished = {Mara Ellison Voss},
  year = {2025},
  url = {https://mouwumou.github.io/Offprint/blog/reading-list-spring/}
}

Voss, M. E. (2025). Three papers that changed how I read data this spring. In Mara Ellison Voss.

Voss, Mara Ellison. “Three Papers That Changed How I Read Data This Spring.” Mara Ellison Voss, 2025.

Voss, Mara Ellison. 2025. “Three Papers That Changed How I Read Data This Spring.” In Mara Ellison Voss. Preprint.

I keep a running note of papers that do not just add a fact but change the lens. Three earned their place this spring.

1. On the intrinsic dimension of representations#

The headline result — that useful representations often live on a surprisingly low-dimensional manifold — is not new. What is new here is the estimator: robust, nearly hyperparameter-free, and honest about its confidence intervals. I have already swapped it into my own analysis.

The practical lesson: report a dimension with an error bar, or do not report one at all.

2. Double descent, revisited without the mystique#

A patient, deflationary paper. It takes the double-descent curve and shows how much of it dissolves once you account for the effective number of parameters rather than the raw count. Not every surprising curve needs a surprising theory; sometimes it needs a better $x$-axis.

3. The GLM that would not die#

A generalized linear model, fit carefully, matched a much larger network on a neural prediction benchmark — and told you why. This is the paper I hand to students who think interpretability and performance are always in tension. They are not; they are often just in different notation.

---

None of these are flashy. All three made my next month of work sharper. That is the only test I trust for a reading list.