Unstable Features, Reproducible Subspaces: Understanding Seed Dependence in Sparse Autoencoders

• URL: http://arxiv.org/abs/2606.12138v1
• Date: Wed, 10 Jun 2026 14:32:57 GMT
• Title: Unstable Features, Reproducible Subspaces: Understanding Seed Dependence in Sparse Autoencoders
• Abstract summary: Sparse autoencoders (SAEs) are widely used to interpret neural network representations.<n>We study this question through emphfeature stability: for each SAE feature, we estimate the probability that a similar feature reappears in an independently trained SAE.<n>We find a pronounced functional asymmetry: stable features carry most of the reconstruction- and prediction-relevant signal, while unstable features have weak marginal impact.
• Score: 7.5916620727820225
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Sparse autoencoders (SAEs) are widely used to interpret neural network representations, but their utility depends on whether the learned features are reproducible across training runs. We study this question through \emph{feature stability}: for each SAE feature, we estimate the probability that a similar feature reappears in an independently trained SAE. This yields a scalable per-feature signal that separates stable from unstable features. In a large-scale study across seeds, models, layers, dictionary sizes, and SAE variants, we find a pronounced functional asymmetry: stable features carry most of the reconstruction- and prediction-relevant signal, while unstable features have weak marginal impact and are dominated by low-frequency surface-form triggers in both activation statistics and automatic explanations. Geometrically, unstable features are individually non-reproducible but concentrate in reproducible lower-rank subspaces, suggesting that seed dependence often reflects basis ambiguity within a shared region of activation space rather than pure noise. A controlled synthetic model makes this mechanism explicit, showing that low-rank ground-truth features can be recovered at the subspace level while remaining non-identifiable as individual SAE latents across seeds. Finally, by pooling unique cross-seed features, we construct more stable SAEs while preserving explained variance in this setting. Together, these results show that unstable features are not merely failed or noisy latents: they have weak individual functional impact, but reflect reproducible low-dimensional structure that standard SAEs resolve differently across seeds.

Related papers

• Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects [58.98080129016082]
  Feature-Effect Geometry Analysis (FEGA) is an unsupervised framework that removes the same active SAE feature across contexts.<n>We show that a feature can be interpretable and causally relevant without providing a stable direction for steering.
  arXiv  Detail & Related papers  (2026-07-27T16:45:08Z)
• C$^{2}$R: Cross-sample Consistency Regularization Mitigates Feature Splitting and Absorption in Sparse Autoencoders [56.67585330425431]
  Sparse Autoencoders (SAEs) are widely used to interpret large language models by decomposing activations into sparse, human-understandable features.<n>These issues stem from inconsistent latent assignment across samples.<n>We introduce C$2$R, which penalizes the co-activation of directionally similar latents.
  arXiv  Detail & Related papers  (2026-06-29T17:45:31Z)
• From Sparse Features to Trustworthy Proxies: Certifying SAE-Based Interpretability [53.681678236115836]
  We study when can an SAE-based explanation be treated as a faithful view of an underlying frozen LM.<n>Our framework derives an upper bound on the base model's expected risk using four measurable quantities.<n>We show that the bound becomes non-vacuous on GPT-2 Small, Gemma-2B, and Llama-3-8B at practical sample sizes.
  arXiv  Detail & Related papers  (2026-06-16T18:28:23Z)
• Preference Instability in Reward Models: Detection and Mitigation via Sparse Autoencoders [20.13794189010598]
  Preference learning in large language models relies on reward models as proxies for human judgment.<n>We analyze this instability at the representation level under three semantic-preserving perturbation types.<n>We propose two SAE-based instability mitigation strategies.
  arXiv  Detail & Related papers  (2026-05-07T16:48:48Z)
• Stability as a Liability:Systematic Breakdown of Linguistic Structure in LLMs [5.96875296117642]
  We show that stable parameter trajectories lead stationary solutions to minimize the forward KL divergence to the empirical distribution.<n>We empirically validate this effect using a controlled feedback-based training framework.<n>It indicates that optimization stability and generative expressivity are not inherently aligned, and that stability alone is an insufficient indicator of generative quality.
  arXiv  Detail & Related papers  (2026-01-26T15:34:50Z)
• Analysis of Variational Sparse Autoencoders [1.675385127117872]
  We investigate whether incorporating variational methods into SAE architectures can improve feature organization and interpretability.<n>We introduce the Variational Sparse Autoencoder (vSAE), which replaces deterministic ReLU gating with sampling from learned Gaussian posteriors.<n>Our findings suggest that naive application of variational methods to SAEs does not improve feature organization or interpretability.
  arXiv  Detail & Related papers  (2025-09-26T23:09:56Z)
• Dense SAE Latents Are Features, Not Bugs [86.50389855919292]
  We show that dense latents serve functional roles in language model computation.<n>We identify classes tied to position tracking, context binding, entropy regulation, letter-specific output signals, part-of-speech, and principal component reconstruction.
  arXiv  Detail & Related papers  (2025-06-18T17:59:35Z)
• Stable Neighbor Denoising for Source-free Domain Adaptive Segmentation [91.83820250747935]
  Pseudo-label noise is mainly contained in unstable samples in which predictions of most pixels undergo significant variations during self-training. We introduce the Stable Neighbor Denoising (SND) approach, which effectively discovers highly correlated stable and unstable samples. SND consistently outperforms state-of-the-art methods in various SFUDA semantic segmentation settings.
  arXiv  Detail & Related papers  (2024-06-10T21:44:52Z)
• Spuriosity Didn't Kill the Classifier: Using Invariant Predictions to Harness Spurious Features [19.312258609611686]
  Stable Feature Boosting (SFB) is an algorithm for learning a predictor that separates stable and conditionally-independent unstable features. We show that SFB can learn anally-optimal predictor without test-domain labels. Empirically, we demonstrate the effectiveness of SFB on real and synthetic data.
  arXiv  Detail & Related papers  (2023-07-19T12:15:06Z)
• Posterior Collapse and Latent Variable Non-identifiability [54.842098835445]
  We propose a class of latent-identifiable variational autoencoders, deep generative models which enforce identifiability without sacrificing flexibility. Across synthetic and real datasets, latent-identifiable variational autoencoders outperform existing methods in mitigating posterior collapse and providing meaningful representations of the data.
  arXiv  Detail & Related papers  (2023-01-02T06:16:56Z)
• Learning Conditional Invariance through Cycle Consistency [60.85059977904014]
  We propose a novel approach to identify meaningful and independent factors of variation in a dataset. Our method involves two separate latent subspaces for the target property and the remaining input information. We demonstrate on synthetic and molecular data that our approach identifies more meaningful factors which lead to sparser and more interpretable models.
  arXiv  Detail & Related papers  (2021-11-25T17:33:12Z)
• Removing Spurious Features can Hurt Accuracy and Affect Groups Disproportionately [83.68135652247496]
  A natural remedy is to remove spurious features from the model. We show that removal of spurious features can decrease accuracy due to inductive biases. We also show that robust self-training can remove spurious features without affecting the overall accuracy.
  arXiv  Detail & Related papers  (2020-12-07T23:08:59Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.