AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

On the Origin of Neural Scaling Laws: from Random Graphs to Natural Language

Maissam Barkeshli, Alberto Alfarano, Andrey Gromov

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Neural scaling laws · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace unbiased walk on ER graph (kappa=0) by power-law-biased transitions (kappa=1) (ablation)2-layer transformer on ER 8K/50K · ER 8K/50Kpower-law fit quality and exponents: alpha_D 1.028 -> 0.925Improved
include the irreducible loss offset E_D in the L(N)_D fit (ablation)Pythia scaling fit · Pythia modelsscaling exponent and fit MSE: nearly a factor of two in exponent, almost an order of magnitude in MSEImproved
use neural network or kernel regression instead of 2d Chinchilla formulaL(N,D) fit · Chinchilla datavalidation MSE: almost 2x improvement in average validation MSEImproved
use 1d fit L(D)_N instead of 2d Chinchilla fitL(D)_N fit · Chinchilla dataMSE: 4.77 x 10^-4 -> 7.76 x 10^-5Improved
use muP instead of standard parameterization2-layer transformer on Fineweb-edu · Fineweb-edumean alpha_D: 0.354 -> 0.501Improved
include embedding parameters in N (ablation)2-layer transformer SP · Fineweb-educompute-optimal exponents: (a,b) = (0.742, 0.386) when excluded; a/b approx 0.94 when includedMixed
increase data complexity from language bigram to T1L2-layer transformer · bigram vs T1Lmean alpha_D: 0.977 -> 0.550Hurt
increase data complexity from T1L to T2L2-layer transformer · T1L vs T2Lmean alpha_D: 0.550 -> 0.497Hurt
use 4-layer instead of 2-layer transformer on T-n-L datatransformer · T2L vs T4Lscaling resultsNo clear effect
use 4 layers instead of 2 layersmuP transformer on Fineweb-edu · Fineweb-edumean alpha_D: 0.501 -> 0.506No clear effect
replace power law fit by exponential fit (ablation)1d scaling fits · ER 8K/50KMSE of L(N)_D fit: 1.07 x 10^-8 -> 7.03 x 10^-8Hurt
add power-law structure via Barabasi-Albert graphER vs BA random walk · random graphsL(N)_D power-law fit qualityHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy