cs.LGNov 12, 2026

When can an interpretability claim be trusted? Stability and certification of feature dimensionality, ridge leverage and the feature recovery rate

Authors: Ali Tajfar, Behnam Tajfar

Organizations: Independent researchers

Abstract

Interpretability tools increasingly support safety claims, such as that a sparse dictionary has recovered most of a model's features. Such claims rest on measurements with arbitrary thresholds, and small changes to the model weights, a different training seed or a different threshold can reverse them. We study two widely used measurements in toy models and synthetic dictionaries with known ground truth: the feature dimensionality of Elhage et al. and the feature recovery rate (FRR) with a cosine threshold. We characterize exactly where feature dimensionality is discontinuous and replace it with a regularized (ridge) version with sharp stability bounds: the ridge leverage score changes by at most 33/(8λ)3\sqrt{3}/(8\sqrtλ) times the operator-norm change of the weights, and this constant cannot be improved; for the regularized dimensionality the stability constant grows like ρ/4ρ/4 in the scale ρ=R/λρ=R/\sqrtλ, with explicit upper and lower bounds. We also prove certificates stating when a claim "feature ii is recovered" must survive a given change of the dictionary. The certificates held in every test and are informative for small perturbations, but certify no feature non-trivially across independently trained dictionaries; at the standard threshold, whether a feature counts as recovered changed with the training seed for 55 percent of features. Experiments use toy models only. The proofs, code and text were drafted by AI agents directed by the authors and checked by an independent blind AI reviewer and by numerical tests; the authors take full responsibility for the content.

Figures & tables

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Jun 16, 2026

From Sparse Features to Trustworthy Proxies: Certifying SAE-Based Interpretability

Sparse autoencoders (SAEs) are increasingly used to extract interpretable features from language models (LMs), yet a central question remains: when can an SAE-based explanation be treated as a faithful view of an underlying frozen LM We study this through a post-hoc generalization framework that certifies the LM via a sparse proxy, obtained by replacing a native hidden activation with its pretrained SAE reconstruction. Our framework derives an upper bound on the base model's expected risk using four measurable quantities: proxy risk, SAE reconstruction gap, concept-pool mismatch, and sparse complexity. We interpret this certificate as an operational criterion for explanatory faithfulness. In particular, a non-vacuous bound indicates that the extracted sparse features retain meaningful predictive information, while small reconstruction and mismatch errors indicate that the proxy remains behaviorally close to the original model. Empirically, we show that the bound becomes non-vacuous on GPT-2 Small, Gemma-2B, and Llama-3-8B at practical sample sizes. A detailed layerwise analysis of Llama-3-8B reveals a strong depth dependence, with later layers becoming much easier to certify, associated with both stronger local fidelity and weaker downstream error amplification. Finally, through feature-shuffling ablations, we show that the decomposition distinguishes genuine semantic alignment from mere statistical sparsity, providing a useful diagnostic for when SAE-based explanations become less reliable.
Sep 29, 2026

Active Budget Can Kill Sensitivity: Diagnosing and Repairing TopK Sparse Autoencoder Reliability

Sparse autoencoders (SAEs) are increasingly scaled to wider dictionaries to recover fine-grained structure from large language model activations. However, a feature is useful for interpretation only if it remains a stable unit of analysis when the same meaning is expressed in different surface forms. We study this reliability question for TopK SAEs via feature sensitivity. Experiments demonstrate that scaling selectively reduces the sensitivity of rare features, while common features remain comparatively stable. A controlled width×k\times k factorial experiment identifies the active budget k as the root cause: the degradation arises from the selection boundary rather than dictionary width alone. We attribute this failure to the geometry of TopK selection. The active margin, the distance to the cutoff, predicts feature loss without thresholds. Guided by this margin diagnosis, we introduce pairwise rank stabilization. Our method targets ordering failures at the cutoff and improves rare-feature sensitivity by 8.838.83 percentage points, while keeping reconstruction and alive-feature coverage near the baseline. Overall, our results suggest that wide TopK SAEs should be evaluated not only by reconstruction, sparsity, and feature count, but also by feature reliability under semantic variation and boundary geometry for stable interpretability.
Oct 30, 2026

Certifying Retraining Regret through Residual Structure

We certify a fixed ridge predictor's regret against full-feature retraining over every continuous completion of an interval-valued table. Our main theorem gives verifiable original-row residual conditions for a uniform error bound independent of nuisance-feature scale and rank, with an asymptotically sharp leading constant and quantitative leakage stability. Full-rank constructions separate certification dimension from feature dimension, including examples requiring directions beyond the candidate. Observed-row anchors make the conditions computable, while exact row geometry and global search return valid upper bounds and feasible original-table witnesses. A finite-scale contraction and common-domain moment fallback implement the structural result. Four real-world tasks retain all model features and certify twenty rank-three instances to width 10−510^{-5}. Their primary median widths improve over centered moments. Joint residual refinement additionally certifies all twenty new response-independent masks to the same precision within 76.21 seconds. For a concrete fixed-point predictor contract, pure energy certifies thirteen of fifteen stopping decisions in ten seconds, versus five for moments. Refinement certifies all fifteen within 91.68 seconds at the same objective threshold.