Research map / Architectures and efficiency / RNNs and normalization
The Quantization Model of Neural Scaling
Eric J Michaud, Ziming Liu, Uzay Girit, Max Tegmark
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 |
|---|---|---|---|
| Scale network parameters and training data | ReLU MLP on multitask sparse parity · Multitask sparse parity | mean test loss | Improved |
| Learn more frequent subtasks before less frequent ones | ReLU MLP on multitask sparse parity · Multitask sparse parity | subtask loss | Improved |
| Increase model size across the first six Pythia models | Pythia · The Pile test set | mean cross-entropy loss: α_N = 0.083 | Improved |
| Increase model size for most language-model tokens | Pythia · The Pile test set | per-token loss | Improved |
| Use QDG gradient clustering to group prediction samples | Pythia 19m · The Pile test set | cluster coherence | Improved |
| Estimate quantum frequency from QDG cluster sizes | Pythia 19m · The Pile test set, QDG clusters | rank-frequency slope: ≈ -1.24; -1.08 | Mixed |
| Use QDG to recover the known subtask frequency distribution | ReLU MLP on multitask sparse parity · Multitask sparse parity subtasks | rank-frequency slope: -1.4; ≈ -1.1 | Hurt |
| Train on a larger dataset in the multi-epoch regime | ReLU MLP on multitask sparse parity · Multitask sparse parity | number of learned subtasks: n ∝ D^{1/(α+1)} | Improved |
| Increase training steps | ReLU MLP on multitask sparse parity · Multitask sparse parity | loss scaling exponent: α_S ≈ 0.29; α_S ≈ 0.45 | Mixed |
| Use spectral clustering with different cluster counts and toy gradient dimensions/noise (ablation) | QDG · QDG toy model | rank-frequency exponent estimate: error is on the order of 0.2 | Mixed |
| Use single-epoch step scaling instead of the theoretical convergence-time assumption (ablation) | ReLU MLP on multitask sparse parity · Multitask sparse parity | subtask convergence time: S_k ∝ p_k^{-0.81}; S_k ∝ p_k^{-1} | Mixed |
Closest papers by meaning
- A Hitchhiker's Guide to Scaling Law Estimation (ICML 2025 · similarity 0.90)
- Language models scale reliably with over-training and on downstream tasks (ICLR 2025 · similarity 0.90)
- Scaling Laws for Precision (ICLR 2025 · similarity 0.90)
- A Simple Model of Inference Scaling Laws (ICML 2025 · similarity 0.90)
- Scaling Law with Learning Rate Annealing (NeurIPS 2025 · similarity 0.89)
- On the Invariance and Generality of Neural Scaling Laws (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.
