Feature Hedging: Correlated Features Break Narrow Sparse Autoencoders

David Chanin, Tomáš Dulka, Adrià Garriga-AlonsoSubmission to ICLR 2026 · Rejected

TL;DR

When a Sparse Autoencoder (SAE) is narrower than the true number of underlying features and those features are correlated, Mean Squared Error (MSE) reconstruction loss forces SAE latents to blend correlated features together. This phenomenon—termed feature hedging—destroys monosemanticity, worsens in narrower models, and explains why techniques like Matryoshka SAEs trade off feature absorption for increased hedging.

What Problem Does This Paper Address?

Sparse Autoencoders (SAEs) are widely used in mechanistic interpretability to decompose dense, polysemantic LLM activations into interpretable linear directions ("monosemantic latents"). However, recent benchmarks reveal that unsupervised SAEs consistently underperform supervised baselines like linear probes on concept detection and steering.

Prior work attributed SAE failures largely to feature absorption, where a specific child feature suppresses a general parent feature due to sparsity penalties. This paper uncovers a complementary, opposite failure mode: feature hedging. In real-world language models, the number of latents is almost certainly smaller than the total number of underlying concepts, and features naturally correlate (or anti-correlate). The authors investigate why and how these conditions cause SAEs to merge distinct concepts into polysemantic mixtures.

Methods

  • Theoretical Derivation: Modeling the SAE decoder under MSE loss as an Ordinary Least Squares (OLS) regression problem. The authors prove that when an untracked ("orphan") feature correlates with a tracked feature, the MSE-optimal reconstruction weight for the orphan feature in the tracked latent is non-zero (proportional to covariance and feature magnitude). Furthermore, an optimal linear encoder acts as a matched filter, causing hedging to appear symmetrically in both the encoder and decoder.
  • Toy Generative Models: Synthetic datasets with ground-truth orthogonal features were constructed under four conditions: fully independent, hierarchical, positively correlated, and anti-correlated firing probabilities.
  • Hedging Degree Metric ($h$): To detect hedging in real LLM SAEs, the authors measure how existing decoder directions shift after adding $N$ new latents. If hedging previously occurred, newly introduced latents "pull out" their mixed components from older latents. The metric projects the shift ($\delta_L$) onto the subspace of newly added latents and subtracts the projection onto random unit-norm latents to discount baseline noise.
  • Balance Matryoshka SAE: Standard Matryoshka SAEs penalize reconstruction across nested subsets of latents to prevent absorption, but their narrow inner levels suffer heavily from hedging. The authors propose adding scaling coefficients $\beta_m$ to each level's loss: $$\mathcal{L} = \sum_{m \in \mathcal{M}} \beta_m \left( \|x - \hat{x}_m\|_2^2 + \lambda S_m \right) + \alpha \mathcal{L}_{\text{aux}}$$ Because absorption adds a negative child component to the encoder while hedging adds a positive one, tuning $\beta_m$ allows these opposing forces to partially cancel out.

Datasets

  • Synthetic LRH Toy Data: Synthetic activations generated from orthogonal feature vectors with controlled joint and conditional activation distributions.
  • The Pile (uncopyrighted): 500M tokens used to train SAEs on the activations of Gemma-2-2b (layer 12) and Llama-3.2-1b (layer 7).
  • SAEBench: Used to benchmark feature absorption, feature splitting, Spurious Concept Removal (SCR), Targeted Probe Perturbation (TPP), and $k$-sparse probing.
  • Penn Treebank Parts of Speech (POS): A custom binary classification dataset created from Penn Treebank POS-tagged sentences to probe whether inner levels of Matryoshka SAEs successfully isolate high-frequency, general syntactic features.

Key Insights and Ideas

  • Reconstruction Loss Drives Hedging: Hedging is not caused by sparsity penalties; it is directly induced by MSE loss. Reconstructing multiple correlated features imperfectly yields lower MSE than reconstructing one feature perfectly while ignoring the other.
  • Anti-Correlation Introduces "Negative" Noise: Anti-correlated features cause latents to mix in negative components while the decoder bias compensates with a positive default offset. In language, where many concepts are negatively correlated, this injects widespread directional noise into latents.
  • Width is Not a Neutral Hyperparameter: Hedging is acutely severe in narrow SAEs and decreases as dictionary width grows. In contrast, feature absorption worsens as SAEs become wider.
  • Hedging vs. Absorption Symmetry:
  • Absorption affects encoder and decoder asymmetrically (the parent encoder shuts off in the presence of the child, while the decoder remains unchanged).
  • Hedging affects encoder and decoder symmetrically (both encoder and decoder rotate toward the correlated feature).

Key Results

  • Empirical Hedging in LLMs: Measuring hedging degree across widths from 128 to 65,536 on Gemma-2-2b and Llama-3.2-1b showed hedging spikes dramatically at widths below 4,096. Even at the maximum width of 65,536, hedging remained above zero.
  • BatchTopK vs. L1: BatchTopK SAEs exhibited systematically higher hedging degrees than $L_1$-regularized SAEs, as strong $L_1$ penalties can slightly reduce positive hedging by penalizing the increased $L_0$ activation rate.
  • Balance Matryoshka Gains: On Gemma-2-2b (width 32,768, 5 nested levels), using a geometric scaling multiplier of around 0.75 between consecutive levels achieved superior performance over both standard SAEs ($\text{multiplier}=0$) and standard Matryoshka SAEs ($\text{multiplier}=1$) on Targeted Probe Perturbation (TPP), feature splitting, and $k=1$ sparse probing.

Limitations and Future Work

  • Scale: Real-model validation was limited to 2B-parameter models (Gemma-2-2b and Llama-3.2-1b) and maximum latent widths of 65,536 due to compute constraints.
  • Measurement Cost: Calculating the hedging degree metric requires fine-tuning an SAE with extended latents alongside an unextended control checkpoint for 250M tokens, making evaluation computationally heavy.
  • Imperfect Cancellation: A single scalar multiplier $\beta_m$ per level cannot perfectly balance all features simultaneously when child features have varying conditional probabilities. Future work includes learning latent-specific balancing coefficients.

Summary updated

Abstract

It is assumed that sparse autoencoders (SAEs) decompose polysemantic activations into interpretable linear directions, as long as the activations are composed of sparse linear combinations of underlying features. However, we find that if an SAE is more narrow than the number of underlying "true features" on which it is trained, and there is correlation between features, the SAE will merge components of correlated features together, thus destroying monosemanticity. In LLM SAEs, these two conditions are almost certainly true. This phenomenon, which we call feature hedging, is caused by SAE reconstruction loss, and is more severe the narrower the SAE. In this work, we introduce the problem of feature hedging and study it both theoretically in toy models and empirically in SAEs trained on LLMs. We suspect that feature hedging may be one of the core reasons that SAEs consistently underperform supervised baselines. Finally, we use our understanding of feature hedging to propose an improved variant of matryoshka SAEs. Importantly, our work shows that SAE width is not a neutral hyperparameter: narrower SAEs suffer more from hedging than wider SAEs.
BibTeX
@misc{chanin2025feature,
  author = {David Chanin and Tomáš Dulka and Adrià Garriga-Alonso},
  title = {{Feature Hedging: Correlated Features Break Narrow Sparse Autoencoders}},
  year = {2025},
  eprint = {2505.11756},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.11756},
}
Figure