IC-053In GPT-2 small's l10h0 name mover queries, the io attribute is encoded with higher-magnitude features than the s attribute, and both are causally relevant, but SAEs preferentially learn io features due to the magnitude asymmetry

Aleksandar Makelov, Georg Lange, Neel Nanda

SourceTowards Principled Evaluations of Sparse Autoencoders for Interpretability and Control

The paper measures the L2 norms of supervised feature vectors for io and s names in the queries of the l10h0 name mover head and finds io features have significantly higher norm. Both attributes are causally relevant: ablating either changes the logit difference by approximately 0.5. Despite this, task SAEs consistently find features for almost all io names but never more than 22 s-name features (out of 216 names), across a wide hyperparameter sweep. Surgically reducing the io feature magnitude causes the number of discovered s features to increase monotonically, and a toy model with isotropic random features of matching norms reproduces the effect.

Evidence
interventional
Key metric
Ablating either io or s in l10h0 queries leads to a change of ≈ 0.5 in logit difference. Number of s features with f1 > 0.5 never exceeds 22 (out of 216 names) across the hyperparameter grid.
Caveat
The magnitude difference between io and s features is described as 'small but significant'; the effect is observed at one specific circuit location (l10h0 queries) and may not generalize to all locations.
Model
GPT-2 Small
Concepts
Linear representation, Method artefact
Datasets
IOI dataset / IOI task [eval]
Methods
Activation patching / Activation replacement / Cross-model activation patching (CMap) [primary], F1 score [eval]
Related work
Wang et al. 2022 (IOI circuit) [builds-on]
Related findings
IC-052
Extraction
automatic-extraction