Search papers, labs, and topics across Lattice.
This paper introduces the Shannon Scaling Law, a theoretical framework that models LLM training as information transmission over a noisy channel, drawing from the Shannon-Hartley theorem. By mapping model parameters to channel bandwidth and training tokens to signal power, the framework captures the interaction between learning signal and intrinsic noise, revealing a fundamental Shannon capacity for LLMs. Experiments on Pythia and OLMo2 under perturbations like noise, quantization, and fine-tuning validate the Shannon Scaling Law, demonstrating its superior accuracy in capturing performance degradation compared to classical scaling laws.
LLMs have a fundamental Shannon capacity: scaling model size or data without preserving a sufficient signal-to-noise ratio inevitably degrades performance.
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) inevitably amplifies noise, inducing a transition from monotonic improvement to U-shaped performance degradation. We validate our theory through experiments on Pythia and OLMo2 under perturbations, including Gaussian noise, quantization and supervised fine-tuning on math, QA and code tasks. The Shannon Scaling Law consistently outperforms classical scaling laws and recent perturbation-aware laws, achieving strong R^2 scores and accurately capturing loss basins missed by prior approaches. It also extrapolates: fitted on leq6.9B Pythia models with leq180B tokens, it predicts the unseen 12B model up to 307B tokens at pooled R^2{=}0.847, while monotonic baselines collapse.
