Newest papers

Strength Controls Change How Interventions Compare Across Sparse Autoencoder Codes

Jinchen Zhang, Zheyuan Liu, Yihan Zhu, Meng Jiang · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26489 · cited by 0 works (OpenAlex) · open-access PDF

French

La modification des codes d'auto-codage sparse (SAE) peut modifier à la fois la direction et la magnitude de l'intervention, de sorte que la comparaison des effets d'intervention nécessite un contrôle de force spécifiant ce qui est maintenu égal. La correspondance normale et la calibration de sortie fournissent des contrôles différents, mais si les interventions produisent des effets différents entre les codes après la correspondance de la force reste incertaine. Nous comparons les effets d'intervention sous ces contrôles pour les groupes de caractéristiques fixes dans Gemma et Llama et observons: (1) l'échelonnage en patch complet. Avec un coefficient par entrée, la calibration Kullback-Leibler (KL) de sortie complète laisse des

Machine translation (Meta M2M100); the English original follows.

Abstract

Changing sparse autoencoder (SAE) codes can alter both intervention direction and magnitude, so comparing intervention effects requires a strength control specifying what is held equal. Norm matching and output calibration provide different controls, but whether interventions yield different effects across codes after strength matching remains unclear. We compare intervention effects under these controls for fixed feature groups in Gemma and Llama and observe: (1) Whole-patch scaling. With one coefficient per input, full-output Kullback-Leibler (KL) calibration leaves smaller mean effect gaps from the encoder-based reference intervention than total-norm matching for nonnegative least-squares (NNLS) refits. On new Llama inputs, total-norm matching loses 11 reference transfers, whereas full-output calibration preserves every transfer decision. (2) Positionwise scaling. On a small matched gender-agreement subset, positionwise norm matching leaves smaller observed mean gaps than full-output calibration. (3) Answer exclusion. Compared with unscaled NNLS, answer-excluded calibration reduces gaps in the measured subject-verb agreement comparisons but increases them for gender refits. Full-output KL calibration also reduces larger gaps for norm-matched constructed candidates, although matching coverage is incomplete in several settings. Since calibration uses the evaluated input's outputs, it serves as a diagnostic of residual disagreement.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Matching (statistics), Calibration, Mathematics, Autoencoder, Residual, Norm (philosophy)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange