cs.LGOct 31, 2026

Sparse Autoencoder Attribution Graphs Are Decided at the Margin

Authors: Karl Li

Organizations: Columbia University

Abstract

Attribution graphs predict sparse autoencoder (SAE) ablations to first order on the active features, but a TopK SAE passes only its k largest preactivations, so small changes often swap active features. On five public TopK SAEs, zeroing one feature changes the active set of the next layer's SAE in 74% to 89% of cases. Our margin graph adds the first order change to every downstream preactivation, then selects the k largest again. On Pythia-70M the margin graph cuts the median error of predicted activations, relative to their exact change, from the attribution graph's 0.739 to 0.121, 84% lower. Exact recomputation of 48 interventions runs 3.5 times faster than the margin graph on our GPU. Patched into Pythia-70M with the clean reconstruction error, the margin graph's activations predict the change in the subject verb agreement logit difference with 1.95 times the attribution graph's mean error, on interventions above a fixed floor. Between seeds of our SAEs, which share one auxiliary loss coefficient, sparse feature maps recover 85% of a graph's squared norm on active features but 47% with each graph masked to its own, 38 percentage points less. Unmasking the inactive tenth nearest the TopK threshold recovers 73% of the masking loss, against 11% for a random tenth, so the margin to that threshold decides how ablations change activations and what graphs carry between seeds.

Figures & tables

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

May 7, 2026

From Token Lists to Graph Motifs: Weisfeiler-Lehman Analysis of Sparse Autoencoder Features

Sparse autoencoders (SAEs) have become central to mechanistic interpretability, decomposing transformer activations into monosemantic features. Yet existing analyses characterise features almost exclusively through top-activating token lists or decoder weight vectors, leaving the higher-order co-occurrence structure shared across features largely unexamined. We introduce a graph-structured representation in which each SAE feature is modelled as a token co-occurrence graph: nodes are the tokens most frequent near strong activations, and edges connect pairs that co-occur within local context windows. A custom WL-style, frequency-binned graph kernel then provides a similarity measure over this structural space. Applied as a proof of concept to features from a large SAE trained on GPT-2 Small and probed with a synthetic mixed-domain corpus, our clustering recovers heuristic motif families (punctuation-heavy patterns, language and script clusters, and code-like templates) that are not recovered by clustering on decoder cosine similarity. A token-histogram baseline achieves higher overall purity, so the contribution of the graph view is complementary rather than dominant: it surfaces structural relationships that token-frequency and decoder-weight views alone do not capture. Cluster assignments are stable across graph-construction hyperparameters and random seeds.
Oct 11, 2026

What Do Sparse Autoencoder Feature Discovery Scores Identify? From Activations to Interventions

Sparse autoencoder (SAE) features are typically discovered from how their activations co-vary with behaviour, yet they are often used through interventions such as steering or ablation. We take intervention attribution, the first-order effect of a specified intervention on the model output, as our estimand. For a broad class of response functions, we show that observational feature-discovery evidence generally does not identify this attribution, and derive an exact gradient-based decomposition of the gap between linear-probe predictions and intervention attribution. Across four model-SAE pairs, median probe-to-attribution magnitude ratios range from tens to hundreds, while attributions closely track small-step steering. For single-latent ablations with nonzero effects, gradient predictions match the realized sign in 91-97% of cases, compared with 17-35% for probe-implied changes. Attribution-selected sets also produce larger absolute mean ablation effects on held-out prompts than sets selected by correlation, rank tests or probe coefficients. Together, these results establish intervention attribution as a principled, gradient-estimable target for selecting SAE features for intervention.
Jul 22, 2026

Are Single-Token Sparse Autoencoder Features Causally Necessary? Layer-Depth and SAE-Family Effects

Sparse autoencoder (SAE) features are used to interpret and steer large language models, yet whether a feature's causal role is stable across SAE families remains untested. Single-token features that activate on one vocabulary item provide the diagnostic case where ground truth permits direct comparison. We analyze 3.9M features across six models and three SAE families using zero-ablation at full layer depth. Single-token features cluster 4.7x tighter in decoder space and concentrate in early layers (Layer 0 in GPT2-Small; L0-L4 in Gemma). Ablating them yields Benjamini-Hochberg-significant logit reductions in 178 of 208 full-layer conditions, with depth controlling whether damage cascades downstream or shapes the output directly. Cross-family causal differences exceed within-family scale effects: on the same base model, GemmaScope and BatchTopK features remain causally anchored, while LlamaScope features are locally redundant. The target token's rank recovers to within 2x baseline 96-98% of the time after the same ablation, and a controlled activation-function comparison reverses sign within the same model, leaving training recipe as the residual candidate. Cross-family interpretability claims are therefore sensitive to training methodology, not just activation function or scale.