IC-160Pythia-70m and Gemma-2-2b implement subject-verb agreement across a relative clause via a circuit of number detectors, PP/RC boundary detectors, and verb form promoters, with Gemma-2-2b additionally using NP number trackers

Samuel Marks, Can Rager, Eric J Michaud, Yonatan Belinkov, David Bau, Aaron Mueller

SourceSparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models

The authors discover sparse feature circuits for subject-verb agreement across a relative clause in both models. In both, the circuit first detects the grammatical number of the main subject, then detects the start of a modifying PP or RC, and finally promotes the matching verb inflection at the end of the clause. Gemma-2-2b additionally employs NP number tracker features that remain active on all tokens within a noun phrase of a given number. The circuits for agreement across a relative clause and across a prepositional phrase share over 85% of their features in both models, indicating a largely uniform handling of these syntactically distinct structures.

Evidence
correlational
Key metric
Pythia circuit: 86 nodes, faithfulness 0.21; Gemma circuit: 223 nodes, faithfulness 0.21; RC and PP circuits share over 85% of features
Caveat
The first 1/3 of the circuit (early layers) is excluded from faithfulness evaluation because train and test splits do not contain identical tokens, making early-segment evaluation unreliable.
Model
Pythia Pythia-70m, Gemma 2 2B
Concepts
Linear representation
Methods
Attribution Patching [supporting], Integrated Gradients / Integral of gradients [supporting]
Related work
Finlayson et al. 2021 [builds-on], Olsson et al. 2022 [context], Wang et al. 2023 [context]
Related findings
IC-161, IC-162
Extraction
automatic-extraction