unified-neural-scaling-laws

name	unified-neural-scaling-laws
description	Unified Neural Scaling Laws (UNSL) methodology for modeling and extrapolating deep neural network scaling behaviors across multiple dimensions (parameters, data size, compute, hyperparameters). Use when analyzing or predicting model performance scaling, optimizing resource allocation across dimensions, or extrapolating training/inference costs for large models. Applicable to vision, language, math, and RL tasks.
license	Complete terms in LICENSE.txt
metadata	{"arxiv_id":"2605.26248","published":"2026-05-25","authors":"Ethan Caballero, Priyank Jaini, David Krueger, Irina Rish","tags":["scaling-laws","deep-learning","neural-networks","model-optimization","compute","extrapolation"]}

Unified Neural Scaling Laws

Core Innovation

Unified functional form that accurately models and extrapolates neural network scaling behavior when multiple dimensions vary simultaneously—addressing the limitation of single-dimension scaling laws.

Scaling Dimensions Covered

UNSL simultaneously models how evaluation metrics vary across:

1. Model parameters (N) - Network size/architecture
2. Training dataset size (D) - Number of training examples
3. Training steps (T) - Optimization iterations
4. Inference steps (I) - Compute at evaluation
5. Total compute (C) - FLOPs budget
6. Hyperparameters (H) - Learning rate, batch size, architecture choices

Key Advantages

• Multi-dimensional extrapolation: Unlike Chinchilla scaling (D + N only), UNSL handles 6+ dimensions simultaneously
• Task-agnostic: Works across vision, language, math, and reinforcement learning
• Architecture-agnostic: Valid across different model architectures within same task domain
• Higher accuracy: Significantly better extrapolation than prior scaling law functional forms

Methodology Framework

1. Functional Form Design

UNSL uses a parametric functional form with:

• Power-law terms: For each scaling dimension
• Interaction terms: Capturing cross-dimensional dependencies
• Saturation terms: Modeling performance limits

General structure:

Metric(N, D, T, I, C, H) = f_base + Σ α_i * dim_i^β_i + Σ γ_jk * dim_j^δ_j * dim_k^δ_k + ...

Where:

• f_base: Baseline performance
• α_i, β_i: Individual dimension scaling coefficients
• γ_jk: Interaction coefficients between dimensions j and k

2. Parameter Estimation

Fit UNSL parameters using:

• Multi-grid sampling: Train models at varied combinations of all dimensions
• Regression fitting: Non-linear least squares or Bayesian optimization
• Cross-validation: Validate extrapolation accuracy on held-out scaling regimes

3. Extrapolation & Prediction

Use fitted UNSL to:

• Predict performance at unseen scales (larger N, more compute, etc.)
• Optimize resource allocation (trade-off between dimensions)
• Estimate training costs before running experiments
• Identify optimal hyperparameters for target scale

Practical Applications

Model Development

Scenario: Planning a 100B parameter language model training

• Use UNSL to predict: optimal data size, compute budget, expected performance
• Trade-off analysis: More parameters vs. more data vs. more compute
• Cost estimation: FLOPs required, training duration, hardware needs

Resource Allocation

Scenario: Fixed compute budget C, optimize (N, D, T) allocation

• Solve: max Metric(N, D, T) subject to N * D * T ≈ C
• UNSL provides closed-form or gradient-based optimization
• Compare against Chinchilla-style allocation (which ignores T and hyperparameters)

Architecture Search

Scenario: Choosing between architectures at different scales

• Fit UNSL separately per architecture family
• Extrapolate to compare at larger scales not yet tested
• Identify which architecture scales better along target dimensions

Downstream Transfer

Scenario: Pretrained model → downstream task performance

• Model upstream scaling → downstream transfer as additional dimension
• Predict: downstream performance from upstream training choices
• Optimize: upstream training for downstream efficiency

Implementation Guide

Data Collection

1. Grid sampling: Train models at systematic combinations of (N, D, T, C, H)

   • Minimum: 3-5 values per dimension
   • Coverage: Ensure combinations span target extrapolation regime

2. Metric tracking: Record evaluation metric at each grid point

   • Primary metric: Task-specific (accuracy, perplexity, reward)
   • Secondary: Training stability, convergence speed

3. Normalization: Scale dimensions to comparable ranges

   • Log-scale: N, D, T, C typically span orders of magnitude
   • Normalize: Hyperparameters to [0, 1] or standard units

Fitting Procedure

# Pseudocode for UNSL fitting
from scipy.optimize import curve_fit

def unsl_function(dimensions, params):
    """
    dimensions: array [N, D, T, I, C, H_1, H_2, ...]
    params: fitted coefficients [α_1, β_1, α_2, β_2, ..., γ_jk, ...]
    """
    N, D, T, I, C, *hyperparams = dimensions
    # Compute individual terms
    term_N = params[0] * N**params[1]
    term_D = params[2] * D**params[3]
    term_T = params[4] * T**params[5]
    # Compute interaction terms
    term_ND = params[6] * (N * D)**params[7]
    # Sum all terms + baseline
    return params[-1] + term_N + term_D + term_T + term_ND + ...

# Fit to data
popt, pcov = curve_fit(unsl_function, training_data, metrics)

Extrapolation Validation

1. Hold-out test: Reserve largest-scale data points for validation
2. Metrics: Measure extrapolation error (RMSE, MAPE) on held-out regime
3. Baseline comparison: Compare against Chinchilla, Kaplan, or other scaling laws
4. Confidence bounds: Use covariance matrix pcov for uncertainty estimates

Comparison to Prior Scaling Laws

Approach	Dimensions	Interaction Terms	Extrapolation Accuracy	Task Coverage
Chinchilla (2022)	N, D	None (independent)	Moderate	Language only
Kaplan (2020)	N	None	Low	Language only
UNSL (2026)	N, D, T, I, C, H	Cross-dimensional	High	Vision, Language, Math, RL

Key Findings

1. Interaction terms critical: Cross-dimensional dependencies (N×D, N×T) significantly improve extrapolation
2. Inference scaling: First scaling law to model inference compute (I) as separate dimension
3. Hyperparameter sensitivity: Learning rate and batch size scale non-linearly with model size
4. Task-specific parameters: UNSL coefficients differ across vision/language/math, but functional form remains valid

Limitations & Caveats

• Training cost: Requires multi-grid experiments (expensive for large models)
• Regime validity: Extrapolation accuracy degrades beyond 10× training data scale
• Architecture dependency: Needs separate fitting for fundamentally different architectures (e.g., CNN vs. Transformer)
• Task boundary: Cannot extrapolate across fundamentally different tasks (e.g., vision → language)

When to Use

• Planning large-scale training: Before committing compute budget
• Resource optimization: Fixed budget, want optimal (N, D, T) split
• Architecture comparison: Extrapolate multiple architectures to larger scale
• Downstream prediction: Estimate transfer performance from upstream choices

When NOT to Use

• Small-scale models: Single-dimension scaling sufficient for <1B parameters
• Novel architectures: UNSL needs fitting data; cannot extrapolate to unseen architectures
• Single-task optimization: If only varying one dimension, simpler scaling laws suffice
• Real-time decisions: UNSL fitting requires offline computation

References

• Paper: Caballero et al. "Unified Neural Scaling Laws" (arXiv:2605.26248, May 2026)
• Prior work: Hoffmann et al. "Training Compute-Optimal Large Language Models" (Chinchilla, 2022)
• Prior work: Kaplan et al. "Scaling Laws for Neural Language Models" (2020)

Related Skills

• quantum-scaling-laws - Scaling behavior in quantum neural networks
• model-architecture-search - Optimizing architecture choices at scale
• compute-budget-optimization - Allocating FLOPs across training stages

Activation Keywords

• unified neural scaling
• UNSL
• multi-dimensional scaling
• scaling laws extrapolation
• compute optimization
• training cost prediction
• resource allocation scaling