Scaling Exponents Across Parameterizations and Optimizers
read the original abstract
Robust and effective scaling of models from small to large width typically requires the precise adjustment of many algorithmic and architectural details, such as parameterization and optimizer choices. In this work, we propose a new perspective on parameterization by investigating a key assumption in prior work about the alignment between parameters and data and derive new theoretical results under weaker assumptions and a broader set of optimizers. Our extensive empirical investigation includes tens of thousands of models trained with all combinations of three optimizers, four parameterizations, several alignment assumptions, more than a dozen learning rates, and fourteen model sizes up to 26.8B parameters. We find that the best learning rate scaling prescription would often have been excluded by the assumptions in prior work. Our results show that all parameterizations, not just maximal update parameterization (muP), can achieve hyperparameter transfer; moreover, our novel per-layer learning rate prescription for standard parameterization outperforms muP. Finally, we demonstrate that an overlooked aspect of parameterization, the epsilon parameter in Adam, must be scaled correctly to avoid gradient underflow and propose Adam-atan2, a new numerically stable, scale-invariant version of Adam that eliminates the epsilon hyperparameter entirely.
This paper has not been read by Pith yet.
Forward citations
Cited by 14 Pith papers
-
GQA-{\mu}P: The maximal parameterization update for grouped query attention
Derives μP scalings for GQA via promoted spectral-norm definition of feature learning and a modified norm preserving scaling laws for non-full-rank matrices, with experiments showing learning-rate transfer.
-
How to Scale Mixture-of-Experts: From muP to the Maximally Scale-Stable Parameterization
The authors derive a Maximally Scale-Stable Parameterization (MSSP) for MoE models that achieves robust learning-rate transfer and monotonic performance gains with scale across co-scaling regimes of width, experts, an...
-
Unlearning with Asymmetric Sources: Improved Unlearning-Utility Trade-off with Public Data
Asymmetric Langevin Unlearning uses public data to suppress unlearning noise costs by O(1/n_pub²), enabling practical mass unlearning with preserved utility under distribution mismatch.
-
Unlearning with Asymmetric Sources: Improved Unlearning-Utility Trade-off with Public Data
ALU uses public data to suppress unlearning cost quadratically while characterizing distribution mismatch effects, enabling mass unlearning with maintained utility.
-
Theory of Optimal Learning Rate Schedules and Scaling Laws for a Random Feature Model
In a random feature model, optimal SGD learning-rate schedules are polynomial decay in the easy phase and warmup-stable-decay in the hard phase, outperforming constant or simple power-law schedules and transferring di...
-
Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
A recurrent-depth architecture enables language models to improve reasoning performance by iterating computation in latent space, achieving gains equivalent to much larger models on benchmarks.
-
Improving Neural Network Training by Decoupling the Magnitude and Direction of Weight Vectors
MD Decoupling factorizes weights into fixed-norm directions and learnable per-row/column magnitudes updated at independent rates, improving Adam and Muon training stability and scale transfer without weight decay or warmup.
-
Learning Rate Transfer in Normalized Transformers
νGPT is a modified parameterization of normalized transformers that enables learning rate transfer across width, depth, and token horizon.
-
C-voting: Confidence-Based Test-Time Voting without Explicit Energy Functions
C-voting improves recurrent reasoning models by selecting among multiple latent trajectories the one with highest average top-1 probability, achieving 4.9% better Sudoku-hard accuracy than energy-based voting and outp...
-
Parcae: Scaling Laws For Stable Looped Language Models
Parcae stabilizes looped LLMs via spectral norm constraints on injection parameters, enabling power-law scaling for training FLOPs and saturating exponential scaling at test time that improves quality over fixed-depth...
-
Two-Point Deterministic Equivalence for Stochastic Gradient Dynamics in Linear Models
Derives a novel two-point deterministic equivalence for random matrix resolvents to obtain unified asymptotics for SGD-trained linear regression, kernel regression, and random feature models.
-
GWT: Scalable Optimizer State Compression for Large Language Model Training
GWT projects gradients into wavelet subspaces to compress optimizer states for memory-efficient LLM training while claiming performance parity with full-rank updates.
-
Statistical Properties of Training & Generalization
Neural scaling laws in deep learning interact with physics constraints and inductive biases beyond classical statistics.
-
Statistical Properties of Training & Generalization
Review of neural scaling laws and their relation to constraints and inductive biases when applying machine learning to physics problems.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.