A new paper on arXiv argues that standard deep neural networks are poorly suited to small-sample regression. In such settings, the authors write, models often collapse to low-variance predictions and fail to capture individual differences.
To address this, they propose a pairwise comparison framework that recasts regression as an aggregation problem. Instead of predicting a value directly, the model learns from comparisons between examples, and the final output is assembled from those pairwise signals.
The abstract presents this as a way to increase data efficiency, but it does not include experimental details or benchmarks. The preprint is described as a cross-listed announcement, and the full methodology and results would need to be examined to assess how well the approach works in practice.