IC-052GPT-2 small's IOI circuit activations are linearly decomposable into features for the io, s, and pos attributes, with the l10h0 name mover's attention decomposing into sparse pairwise feature interactions

Aleksandar Makelov, Georg Lange, Neel Nanda

SourceTowards Principled Evaluations of Sparse Autoencoders for Interpretability and Control

The paper computes supervised mean feature dictionaries for the io, s, and pos attributes at every cross-section of the IOI circuit in GPT-2 small and tests whether the reconstructions are sufficient and necessary for the task via activation patching. Both tests pass at high levels for the supervised dictionaries. Additionally, the attention scores of the l10h0 name mover head decompose into pairwise dot products between query and key features, with most interactions near zero and only three significant ones: s-query with s-key, pos-query with pos-key, and a negative io-query with io-key interaction, indicating the head inhibits both names and relies on position to distinguish them.

Evidence
interventional
Key metric
GPT-2 small achieves accuracy of ≈ 91% on the IOI test set; average logit difference ≈ 3.3; correct name has higher logit in ≈ 99% of examples. On the greater-than task, exchanging 2 SAE latents matched next-token predictions 90% of the time (vs 38% for no intervention).
Caveat
The evaluation is restricted to the IOI circuit cross-sections identified by Wang et al. (2022) and to three specific attributes (io, s, pos); the authors note the tasks are 'somewhat contrived/templatic' and that the choice of attributes is a crucial modeling decision.
Model
GPT-2 Small
Concepts
Linear representation, Feature interaction
Datasets
IOI dataset / IOI task [eval], OpenWebText / OpenWebText-10k [source]
Methods
Activation patching / Activation replacement / Cross-model activation patching (CMap) [primary]
Related work
Wang et al. 2022 (IOI circuit) [builds-on], Bricken et al. 2023 (SAE) [context]
Related findings
IC-053
Extraction
automatic-extraction