Stable and Steerable Sparse Autoencoders with Weight Regularization
| Authors | Piotr Jedryszek & Oliver M. Crook |
| Year | 2026 |
| Field | Statistics / ML |
| arXiv | 2603.04198 |
| Download | |
| Categories | stat.ML, cs.LG |
Abstract
Sparse autoencoders (SAEs) are widely used to extract human-interpretable features from neural network activations, but their learned features can vary substantially across random seeds and training choices. To improve stability, we studied weight regularization by adding L1 or L2 penalties on encoder and decoder weights, and evaluate how regularization interacts with common SAE training defaults. On MNIST, we observe that L2 weight regularization produces a core of highly aligned features and, when combined with tied initialization and unit-norm decoder constraints, it dramatically increases cross-seed feature consistency. For TopK SAEs trained on language model activations (Pythia-70M-deduped), adding a small L2 weight penalty increased the fraction of features shared across three random seeds and roughly doubles steering success rates, while leaving the mean of automated interpretability scores essentially unchanged. Finally, in the regularized setting, activation steering success becomes better predicted by auto-interpretability scores, suggesting that regularization can align text-based feature explanations with functional controllability.
Engineering Breakdown
Plain English
This paper addresses a critical practical problem with sparse autoencoders (SAEs): the features they learn vary dramatically depending on random initialization and training hyperparameters, making them unreliable for scientific interpretation. The authors systematically study how adding L1 and L2 weight regularization penalties to the encoder and decoder improves feature stability across different random seeds. On MNIST, they show that L2 regularization combined with tied initialization and unit-norm decoder constraints substantially increases the consistency of learned features across runs. When applied to TopK SAEs trained on language model activations (Pythia-70M), even a small L2 penalty increases the fraction of features that remain consistent across three independent training runs.
Core Technical Contribution
The core insight is that weight regularization—a simple but underexplored technique in SAE training—directly addresses the reproducibility and interpretability crisis in mechanistic interpretability work. The authors provide the first systematic empirical characterization of how L1 and L2 penalties interact with standard SAE architectural choices (tied weights, unit-norm constraints, different sparsity methods like TopK). They demonstrate that regularization doesn't just improve generalization; it fundamentally changes the optimization landscape to converge toward more stable, redundant feature representations that are robust to initialization noise. This is novel because prior SAE work treated regularization as optional or incidental, when in fact it can be a primary lever for reproducible interpretability research.
How It Works
Sparse autoencoders consist of an encoder that maps neural network activations to a sparse latent code, and a decoder that reconstructs the activation. The standard training objective minimizes reconstruction loss plus a sparsity penalty (e.g., L0 or TopK penalty). The authors augment this with L1 or L2 regularization on the encoder weights W_enc and decoder weights W_dec, adding terms like λ||W_enc||_2² or λ||W_dec||_1 to the loss function. In addition, they combine regularization with architectural constraints: tied weights (encoder and decoder share parameters), unit-norm decoder columns (each decoder feature has ||d_i|| = 1), and different sparsity mechanisms. The key mechanism is that L2 regularization pushes weights toward small magnitudes uniformly, which prevents any single feature from dominating the optimization landscape; this creates a 'core' of stable features that emerge robustly across seeds. For TopK sparsity, the authors show that regularization interacts favorably because it dampens the dominance of spurious features that only appear in high-seed runs, allowing the true structure in the data to be discovered more consistently.
Production Impact
For teams building mechanistic interpretability pipelines—whether for safety audits, model debugging, or feature extraction—this work directly improves the reliability of SAE-based analysis. Previously, engineers had to train multiple SAEs and manually merge findings across seeds, or accept that feature discovery was inherently noisy. With regularization, a single SAE training run produces more reproducible features that can be confidently published or used in downstream analysis. The implementation cost is negligible: add one hyperparameter (regularization strength λ), which typically requires minimal tuning (the paper suggests small values like 0.001–0.01 work broadly). However, there is a mild trade-off: regularization slightly increases reconstruction loss on the training distribution, so practitioners must validate that the loss remains acceptable for their downstream task (e.g., feature reconstruction for interpretability explanations). For language model auditing workflows, this could reduce the need for ensemble SAEs by 2–3×, directly lowering GPU compute and wall-clock training time.
Limitations and When Not to Use This
The paper's evaluation is limited to relatively small models (MNIST and Pythia-70M) and does not include large-scale language models (Llama, GPT-scale), leaving open whether stability gains hold at 7B+ parameters where feature geometry may be qualitatively different. The mechanism behind why regularization produces stability is empirically demonstrated but not theoretically explained—the authors don't provide a loss landscape analysis or generalization bounds that would predict regularization strength from data properties. Additionally, the paper assumes that 'stable features' are the same as 'true features': it's possible that regularization is simply biasing toward a particular basin of attraction that is reproducible but not semantically meaningful (a form of implicit inductive bias that helps reproducibility at the cost of interpretability). The paper also doesn't address how regularization interacts with other SAE variants (e.g., gated SAEs, SAEs with auxiliary losses) or whether the findings transfer to different activation types (e.g., attention head outputs vs. MLP activations).
Research Context
This work builds directly on the mechanistic interpretability pipeline developed by Anthropic and others, which uses SAEs to decompose neural network activations into interpretable features (prior work: Sharkey et al., Bau et al.). It responds to a growing frustration in the community that SAE features lack reproducibility—a problem well-documented in recent papers on SAE instability and the difficulty of validating feature semantics. The contribution is incremental but practically important: rather than proposing a new SAE architecture, it shows that careful tuning of a classical regularization technique can solve a real failure mode. This opens a research direction into the interaction between inductive biases (regularization, constraints, initialization) and feature stability, which could lead to principled guidelines for training interpretable models. The work also implicit criticism of the field's focus on novel architectures over careful empirical engineering.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::