A paper by Haotian Ma, Ziye Mei and Frederic Sala, published on October 15 and submitted to Hugging Face on October 16, finds that sparse autoencoder (SAE) features can describe model behavior more compositionally than matched random rotations, but accurate explanations and interventions still require modeling interactions between features. The authors evaluated 18 model and behavior settings.

The researchers introduce SAE Additivity Residuals (SAR), a measure of behavioral variation left unexplained when predictions include interactions only up to a chosen order. Under independent feature subset sampling, population SAR equals the normalized energy of interactions above that order and bounds coalition ordering error for the best low-order approximation. Coalition ordering compares groups of features by their predicted behavioral effects.

To separate compositionality in the learned SAE representation from effects of later computation, the team compared learned coordinates with sum-preserving random rotations of the same selected subspace. Learned coordinates had lower first-order residuals across all 18 settings, but were not fully additive. Pairwise interaction terms improved held-out coalition ordering in every setting and beat parameter-matched models with randomly selected third-order terms.

In a combined steering experiment, pairwise modeling raised rank fidelity from 0.5874 to 0.7790 and reduced ordering error from 0.1444 to 0.0606. A separate paper summary on the page reports that pairwise modeling raised held-out coalition rank fidelity from 0.9782 to 0.9959. The authors say the findings support using SAE features to describe behavior while accounting for interactions in joint explanations and interventions. They released code and a visual five-minute tutorial through OpenSAE.

A paper by Haotian Ma, Ziye Mei and Frederic Sala, published on October 15 and submitted to Hugging Face on October 16, finds that sparse autoencoder (SAE) features can describe model behavior more compositionally than matched random rotations, but accurate explanations and interventions still require modeling interactions between features. The authors evaluated 18 model and behavior settings.

The researchers introduce SAE Additivity Residuals (SAR), a measure of behavioral variation left unexplained when predictions include interactions only up to a chosen order. Under independent feature subset sampling, population SAR equals the normalized energy of interactions above that order and bounds coalition ordering error for the best low-order approximation. Coalition ordering compares groups of features by their predicted behavioral effects.

To separate compositionality in the learned SAE representation from effects of later computation, the team compared learned coordinates with sum-preserving random rotations of the same selected subspace. Learned coordinates had lower first-order residuals across all 18 settings, but were not fully additive. Pairwise interaction terms improved held-out coalition ordering in every setting and beat parameter-matched models with randomly selected third-order terms.

In a combined steering experiment, pairwise modeling raised rank fidelity from 0.5874 to 0.7790 and reduced ordering error from 0.1444 to 0.0606. A separate paper summary on the page reports that pairwise modeling raised held-out coalition rank fidelity from 0.9782 to 0.9959. The authors say the findings support using SAE features to describe behavior while accounting for interactions in joint explanations and interventions. They released code and a visual five-minute tutorial through OpenSAE.