When one dominant signal distorts sparse autoencoder scores

When one direction dominates

Sparse autoencoders are trained to represent model activity using a sparse set of features. Two common ways to assess them are to measure how much activity they reconstruct and to measure how much the model is affected when features are removed, or ablated.

In Massive Activations Confound Sparse Autoencoders, Nathan Delisle and Chenhao Tan argue that these evaluations can be distorted when activity is concentrated in the first principal component (PC1)—the direction that accounts for the most variance in a dataset. The abstract reports that, in some cases, up to 99.9% of activation variance lies along PC1.

The authors say PC1 has no unique sparse decomposition: different SAE training runs can represent it using different feature sets. As a result, features that rank highly by importance may not be shared across runs. The abstract reports that an average of 47 of the 50 most important features in one run were unmatched in a separate run, compared with about 27 unmatched features in a randomly chosen set of 50.

The abstract also gives an example of a dictionary reporting 99.96% of variance explained while leaving 33% of the variance outside PC1 unexplained. This illustrates the authors’ concern that a high overall reconstruction score can obscure performance on activity beyond the dominant component.

Separating PC1 during training

The proposed correction is to remove PC1 from activations before training the SAE, then add PC1 back to the SAE’s reconstruction. The stated aim is to keep the dominant component from being represented through run-dependent sparse features while still including it in the reconstruction.

For affected models, the authors report that the rate of shared features among the top 50 rose from 4–39.5% to 73–99.5% with this approach. At the same TopK sparsity, the corrected SAE reportedly degraded the model up to 44% less and reduced content reconstruction error by up to 33.9%.

What remains uncertain

These results suggest that PC1 can affect feature rankings and evaluation scores in the cases studied. They do not establish that all SAE evaluations are affected, or that the proposed correction will help every model. The abstract does not identify the models or provide details needed to assess how results vary across architectures, layers, datasets, and SAE configurations. The “up to” figures describe maximum reported improvements, not typical gains.

This account is based on the abstract alone; the supplied evidence does not include the paper’s full methods, tables, or discussion of limitations. The work is an arXiv preprint, so its claims should be regarded as provisional. The abstract describes a correction and quantitative results, but is not enough to judge how broadly they generalize.