AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Deriving Neural Scaling Laws from the Statistics of Natural Language

Francesco Cagnetta, Allan Raventos, Surya Ganguli, Matthieu Wyart

ICML 2026 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data · 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
rescale n-gram losses by n^-gamma and data by n^(2 beta)n-gram learning curves of GPT-2 style transformers · TinyStories and WikiTextcurve collapseImproved
vary architecture (GPT-2 APE, RoPE, LLaMA, Mamba, n-gram) for estimating gamma (ablation)conditional entropy decay exponent · TinyStoriesgammaNo clear effect
increase maximal context length T (ablation)autoregressive loss · TinyStoriespower-law decay with PNo clear effect
take beta outside its standard-error interval (ablation)scaling collapse · TinyStoriescollapse qualityHurt
take gamma outside [0.31, 0.34] (ablation)scaling collapse · TinyStoriescollapse qualityHurt
take gamma outside [0.23, 0.30] (ablation)scaling collapse · WikiTextcollapse qualityHurt
use beta in [0.6, 0.7] instead of measured beta (ablation)LLaMA model · TinyStoriescollapse sharpnessImproved
increase P by 10x beyond threshold; within-horizon excess lossn-gram losses n<=12 · TinyStories GPT-2 T=128decay exponent delta_n: delta_n > gamma/(2beta)Improved

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