A new preprint on arXiv examines the limits of what can be learned about centered quadratic networks from two common diagnostic tools: spectral summaries and scalar learning curves. The authors ask whether these summaries are sufficient to identify the network's structure, a question with implications for interpretability and model analysis.

To probe this, they construct analytic positive input densities that have exactly Gaussian one- and two-coordinate marginals. These carefully designed inputs allow them to isolate how much information the summaries actually carry about the network, rather than about the input distribution.

The abstract is brief and does not spell out the full conclusions, but the framing suggests that identifiability is not guaranteed in general. The paper is available on arXiv under identifier 2610.20678.