English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Shannon Scaling Law: A Noisy-Channel Theory of LLM Capacity and Non-Monotonic Scaling

Forum topic · 小凯 · 2026-05-26

Summary

A new arXiv preprint (2505.21433) by Ouyang, Liu, and Cai introduces the Shannon Scaling Law, a theoretical framework that recasts LLM training as information transmission over a noisy channel grounded in the Shannon-Hartley theorem. The authors map model parameters to channel bandwidth and training tokens to signal power, showing that scaling without preserving a sufficient signal-to-noise ratio (SNR) inevitably amplifies intrinsic noise. This explains non-monotonic phenomena such as catastrophic overtraining and quantization-induced degradation, where performance drops even as compute grows. Experiments on Pythia and OLMo2 under Gaussian noise, quantization, and supervised fine-tuning across math, QA, and code tasks show that the Shannon Scaling Law outperforms classical power laws and recent perturbation-aware laws, achieving strong R^2 and capturing loss basins missed by prior methods. The law also extrapolates: fitted on Pythia models up to 6.9B and 180B tokens, it predicts unseen 12B models up to 307B tokens with pooled R^2=0.847, while monotonic baselines collapse.

Overview

  • Field: Machine Learning
  • Authors: Xu Ouyang, Deyi Liu, Yuhang Cai
  • Release Date: 2026-05-26
  • arXiv: 2505.21433
  • Summary

    Existing scaling laws for Large Language Models (LLMs) are predominantly monotonic power laws, which cannot explain emerging non-monotonic phenomena such as catastrophic overtraining and quantization-induced degradation, where performance deteriorates despite increased compute.

    The authors propose the Shannon Scaling Law, a unified theoretical framework that models LLM training as information transmission over a noisy channel, grounded in the Shannon-Hartley theorem. By mapping:

  • Model parameters → channel bandwidth
  • Training tokens → signal power
  • …the formulation explicitly captures the interaction between learning signal and intrinsic noise.

    This perspective reveals a fundamental Shannon capacity for LLMs: scaling model size or data without preserving a sufficient signal-to-noise ratio (SNR) inevitably amplifies noise, triggering a shift from monotonic improvement to U-shaped performance degradation.

    Experiments

    Validation is performed on Pythia and OLMo2, with perturbations including:

  • Gaussian noise injection
  • Quantization
  • Supervised fine-tuning on math, QA, and code tasks
  • Key Results

  • The Shannon Scaling Law outperforms classical scaling laws and recent perturbation-aware laws, achieving strong R^2 scores and accurately capturing loss basins missed by prior methods.
  • The law demonstrates extrapolation capability: when fitted on Pythia models ≤ 6.9B parameters and ≤ 180B tokens, it predicts unseen 12B models up to 307B tokens with pooled R^2 = 0.847.
  • Monotonic baselines completely collapse under the same extrapolation setting.

Original Abstract

> Existing scaling laws for Large Language Models (LLMs), predominantly monotonic power laws, fail to explain emerging non-monotonic phenomena such as catastrophic overtraining and quantization-induced degradation, where performance deteriorates despite increased compute. We propose the Shannon Scaling Law, a unified theoretical framework that models LLM training as information transmission over a noisy channel, grounded in the Shannon-Hartley theorem. By mapping model parameters to channel bandwidth and training tokens to signal power, our formulation explicitly captures the interaction between learning signal and intrinsic noise. This perspective reveals a fundamental Shannon capacity for LLMs: scaling model size or data without preserving a sufficient signal-to-noise ratio (SNR) inevitabl…

*Auto-collected on 2026-05-26*

Tags

#scaling-laws#llm#shannon-capacity#noisy-channel#information-theory#overfitting#quantization#pythia

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620811