Skip to content

Superposition Yields Robust Neural Scaling

Conference: NeurIPS 2025 arXiv: 2505.10465 Code: GitHub Area: LLM Pre-training Keywords: Neural scaling laws, superposition, representation learning, LLM theory, weight decay

TL;DR

This paper identifies representational superposition as the core driver of neural scaling laws: in the strong-superposition regime, loss universally scales inversely with model dimension (\(L \propto 1/m\)), independent of the specific form of the data frequency distribution—consistent with empirical scaling behavior in real LLMs.

Background & Motivation

Neural scaling laws are a central empirical regularity in modern AI: larger models achieve lower loss, following a power-law relationship. Yet their origin remains poorly understood.

Limitations of existing explanations: - Data-manifold/function-fitting theories require the data distribution itself to be a power law in order to produce power-law scaling - Skill-learning models (Hutter 2021, Michaud et al. 2023) similarly rely on power-law distribution assumptions - Kernel-method analyses depend on power-law decay of eigenvalues - These explanations operate in the weak-superposition regime (finite-variance regime), which may not match the regime in which LLMs actually operate

Key observation: LLMs must represent over fifty thousand tokens and numerous abstract concepts within hidden spaces of only a few thousand dimensions. This implies that LLMs necessarily operate in a superposition regime—representing far more features than the model dimension.

Core Problem: How does superposition affect neural scaling laws?

Method

Overall Architecture

The paper adopts Anthropic's (2022) superposition toy model (an autoencoder) and systematically studies how the degree of superposition affects scaling behavior.

Input generation: \(x_i = u_i v_i\), where \(u_i \sim \text{Bernoulli}(p_i)\), \(v_i \sim U(0,2)\) - \(p_i\) is the frequency (importance) of feature \(i\), sorted by frequency - Activation density \(E = \sum_i p_i\)

Model: \(h = W^T x\) (encoding), \(y = \text{ReLU}(Wh + b)\) (decoding) - \(W \in \mathbb{R}^{n \times m}\), \(n\) is the number of features, \(m\) is the model dimension, \(m \ll n\) - Loss \(L = \langle \|y - x\|_2^2 \rangle_x\)

Key Designs: Controlling Superposition via Weight Decay

The paper introduces decoupled weight decay (positive or negative) to systematically control the degree of superposition:

\[W_{i,t+1} = \begin{cases} W_{i,t} - \eta_t \gamma W_{i,t}, & \gamma \geq 0 \\ W_{i,t} - \eta_t \gamma W_{i,t}(1/\|W_{i,t}\|_2 - 1), & \gamma < 0 \end{cases}\]
  • \(\gamma > 0\) (positive weight decay): suppresses superposition → weak-superposition regime
  • \(\gamma < 0\) (negative weight decay): encourages unit norms → strong-superposition regime

Superposition metric: \(\phi_{1/2} = |\{i: \|W_i\|_2 > 1/2\}| / n\) - Weak superposition: \(\phi_{1/2} \approx m/n\) (only the \(m\) most important features are represented) - Strong superposition: \(\phi_{1/2} \approx 1\) (nearly all features have representations)

Analysis in the Weak-Superposition Regime

Under ideal non-superposition conditions, the top \(\phi_{1/2} n\) most frequent features are perfectly represented and the rest are ignored:

\[L \approx \langle v^2 \rangle \sum_{i > \phi_{1/2} n} p_i\]

When \(p_i \propto 1/i^\alpha\), \(L \propto m^{-(\alpha-1)}\) (a power law only when \(\alpha > 1\)).

Conclusion: Under weak superposition, the existence and exponent of the scaling law depend on the specific form of the data frequency distribution.

Analysis in the Strong-Superposition Regime

The dominant loss source becomes geometric overlap between representation vectors \((W_i \cdot W_j)^2\).

Key geometric properties: 1. Random unit vectors: The mean squared inner product of two random unit vectors in \(\mathbb{R}^m\) is \(1/m\) 2. Equiangular tight frames (ETF): Representations of approximately \(m^2/2\) important features approach an ETF structure 3. Welch bound: \(\max_{i \neq j} |w_i \cdot w_j| \geq \sqrt{\frac{\nu - m}{m(\nu - 1)}} \approx 1/\sqrt{m}\)

Consequently, squared overlap typically scales as \(1/m\), yielding:

\[L \propto 1/m \quad (\alpha_m = 1)\]

When feature frequencies are more skewed (large \(\alpha\)), the ETF-like feature contribution becomes negligible, leading to \(\alpha_m \approx 2(\alpha - 1)\).

Loss & Training

AdamW optimizer with warmup and cosine decay learning rate schedule. New data is sampled at each step. \(n = 1000\) is fixed; \(m\) varies from 10 to 100.

Key Experimental Results

Main Results: Toy Model

Regime Data exponent \(\alpha\) Model exponent \(\alpha_m\) Power law? Distribution-dependent?
Weak superposition \(\alpha = 0.5\) No power law ✗ ✓
Weak superposition \(\alpha = 1.0\) \(\approx 0\) Marginal ✓
Weak superposition \(\alpha = 2.0\) \(\approx 1.0\) ✓ ✓
Strong superposition \(\alpha = 0.5\) \(\approx 1.0\) ✓ ✗
Strong superposition \(\alpha = 1.0\) \(\approx 1.0\) ✓ ✗
Strong superposition \(\alpha = 2.0\) \(\approx 1.3\) ✓ ✗

Validation on Real LLMs

Analysis of four open-source model families (OPT, GPT-2, Qwen, Pythia):

Observation Result
Mean squared inner product of row-normalized language model head \(W\) Approximately follows \(1/m\) scaling
Loss vs. model dimension \(L = C_m/m^{\alpha_m} + L_{\backslash m}\), \(\alpha_m = 0.91 \pm 0.04\)
Inferred from Chinchilla \(\alpha_m = (2.52 \pm 0.03) \times 0.35 = 0.88 \pm 0.06\)
Whether LLMs operate in superposition ✓ Confirmed (supported by row-norm and interference distributions in the language model head)

Ablation Study

  • Activation density \(E\): Does not affect scaling behavior (verified in Appendix D.4)
  • Weight decay value \(\gamma\): Systematically controls the degree of superposition; small \(\gamma\) → strong superposition, large \(\gamma\) → weak superposition
  • Cross-entropy vs. squared-error loss: Does not affect scaling behavior (demonstrated in Appendix A.2)
  • ETF vs. random vectors: Representations of important features are closer to ETF (lower variance), but the mean is \(1/m\) in both cases

Key Findings

  1. The strong-superposition regime yields robust \(1/m\) scaling, independent of the specific form of the data frequency distribution
  2. Scaling laws in the weak-superposition regime are sensitive to the data distribution—power-law frequencies are necessary to produce power-law scaling
  3. Real LLMs operate in the strong-superposition regime; \(\alpha_m \approx 1\) is consistent with theoretical predictions
  4. Loss decomposes into a model-size-dependent term (representation loss) and a model-size-independent term (intrinsic data uncertainty)

Highlights & Insights

  1. Unified explanation: The origin of scaling laws is attributed to geometry—interference between representation vectors scales as \(\sim 1/m\), yielding an elegant and intuitive account
  2. Robustness finding: In the strong-superposition regime, the scaling exponent is approximately 1 regardless of data distribution details—explaining the universality of neural scaling laws
  3. New role of weight decay: This work is the first to systematically demonstrate that weight decay controls the degree of superposition, with direct implications for practical training
  4. Verifiable prediction: nGPT (which constrains hidden states to the unit hypersphere) encourages superposition and should therefore be more parameter-efficient—preliminary evidence supports this
  5. Theory–experiment loop: A complete chain of reasoning is established, from precise analysis of the toy model to empirical validation on real LLMs

Limitations & Future Work

  1. Lack of rigorous mathematical proof: Analysis of the strong-superposition regime is primarily based on observation and heuristic reasoning; the model is not solved exactly
  2. Only representation loss is analyzed: LLM loss also includes processing loss from Transformer layers \(f_\ell(\ell)\), which is not studied independently
  3. Data/training-step scaling not analyzed: Only model-width scaling is studied; data-size scaling is left for future work
  4. Gap between toy model and LLMs: The toy model lacks Transformer layers, uses a different loss function, and simplifies data structure
  5. Causality not established: \(\alpha_m \approx 1\) in LLMs may have alternative explanations (e.g., depth–width balance)
  • Relation to Kaplan et al. (2020) Chinchilla scaling laws: This paper provides a mechanistic explanation for those laws
  • Relation to Anthropic (2022) superposition model: The framework is directly inherited, but the relationship between superposition and scaling is studied systematically for the first time
  • Relation to Michaud et al. (2023) quantization model: Weak-superposition results are consistent with prior work; strong-superposition results are entirely new
  • Implication: Encouraging superposition (e.g., nGPT, optimizers without weight decay) may be an effective means of improving LLM efficiency

Rating

  • Novelty: ⭐⭐⭐⭐⭐ — The insight that superposition is the core mechanism underlying scaling laws is entirely original
  • Theoretical Depth: ⭐⭐⭐⭐ — The geometric argument is intuitive and compelling, though rigorous proofs are absent
  • Experimental Thoroughness: ⭐⭐⭐⭐ — Toy-model experiments are comprehensive and LLM validation is solid, but intervention experiments are lacking
  • Writing Quality: ⭐⭐⭐⭐⭐ — Well-illustrated, clearly explained, and excellently structured
  • Value: ⭐⭐⭐⭐ — Offers direct guidance for training strategies (weight decay, architecture choices)
  • Overall: ⭐⭐⭐⭐⭐ (9/10) — Elegantly connects two major themes in AI research: mechanistic interpretability (superposition) and scaling laws