Research map / Large language models / Language, safety and interpretability
LLMs as Noisy Channels: A Shannon Perspective on Model Capacity and Scaling Laws
Xu Ouyang, Deyi Liu, Yuhang Cai, Jing Liu, Yuan Yang, Chen Zheng, Thomas Hartvigsen, Yiyuan Ma
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use token-dependent noise term c(DN)^gamma instead of size-only cN^gamma (ablation) | Shannon Scaling Law · Pythia 10 dB | R^2: 0.8035 -> 0.9555 | Improved |
| same, OLMo2 at 10 dB (ablation) | Shannon Scaling Law · OLMo2 10 dB | R^2: 0.8695 -> 0.8695 | No clear effect |
| same, Pythia 40 dB (ablation) | Shannon Scaling Law · Pythia 40 dB | R^2: 0.9895 -> 0.9895 | No clear effect |
| use 9-parameter full law rather than 6-parameter simplified for joint extrapolation (ablation) | Shannon · Pythia k=5,j=12 | pooled R^2: 0.673 -> 0.847 | Improved |
| use 6-parameter simplified law for single-axis model extrapolation (ablation) | Shannon · Pythia k=4 | pooled R^2: 0.787 -> 0.837 | Improved |
| use Shannon law instead of Asymmetric law | Scaling law fit · GPTQ 2-bit Pythia | R^2: 0.9499 -> 0.9602 | Improved |
| use Shannon law instead of Chinchilla | Scaling law fit · Gaussian noise Pythia avg | R^2: 0.5289 -> 0.9613 | Improved |
| use Shannon law instead of OpenAI law | Scaling law fit · SFT StarCoder avg | R^2: -1.010 -> 0.937 | Improved |
| use Shannon law instead of Asymmetric Law | Scaling law fit · SFT GSM8K avg | R^2: 0.896 -> 0.936 | Improved |
| use Shannon law instead of Asymmetric Law | Scaling law fit · Gaussian noise OLMo2 10 dB | R^2: 0.8708 -> 0.8695 | Hurt |
| use Shannon-Simpl instead of Shannon 9p | Scaling law extrapolation · Pythia token j=15 | pooled R^2: 0.941 -> 0.945 | No clear effect |
| use Shannon instead of OpenAI | Scaling law · Pythia model extrapolation k=4 | pooled R^2: -0.048 -> 0.787 | Improved |
Closest papers by meaning
- Understanding LLM Behaviors via Compression: Data Generation, Knowledge Acquisition and Scaling Laws (NeurIPS 2025 · similarity 0.89)
- Compute-Optimal LLMs Provably Generalize Better with Scale (ICLR 2025 · similarity 0.89)
- LLMs on the Line: Data Determines Loss-to-Loss Scaling Laws (ICML 2025 · similarity 0.89)
- Universal One-third Time Scaling in Learning Peaked Distributions (ICML 2026 · similarity 0.89)
- Language models scale reliably with over-training and on downstream tasks (ICLR 2025 · similarity 0.89)
- An Empirical Study on Noisy Data and LLM Pretraining Loss Divergence (NeurIPS 2026 · similarity 0.89)
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.
