Hi all, thanks to ARC and AICrowd for running this challenge! It’s been a fantastic problem to work on ![]()
I’m sharing a writeup on what I believe is the core reason cumulant propagation fails at depth, with a proven fix. In a nutshell: when depth grows with width, a collective scale mode builds up that poisons the order-by-order expansion. Separating it out exactly via ReLU homogeneity and carrying it as a one-parameter law restores convergence and beats sampling at every width tested on across networks of various sizes (64x16 through 1024x256, including improving ARC’s MSE on the competition’s own 1024x16, beating both sampling and existing kappa methods with no fitted constants).
PDF and repo with everything needed to reproduce the results are attached. Thanks again!
Pranav Viswanath ARC Writeup.pdf (768.4 KB)