Kolmogorov-Arnold networks (KANs) have been noted for a dimension-independent approximation exponent, provided the target function admits a sufficiently smooth representation by univariate edge functions. That property would be a major advantage over many standard architectures, which typically suffer from the curse of dimensionality.
The new arXiv paper, titled "Kolmogorov-Arnold Networks: Curse of Dimensionality and Structural Bias," asks directly whether this favorable scaling belongs to the architecture itself or to a bias in how the problem is posed. The abstract raises the question but cuts off mid-sentence, so the source does not reveal the authors' answer or evidence.
Because the abstract is incomplete, this article can only report the question, not the outcome. Readers interested in the resolution will need to consult the full paper on arXiv.