Do Attribution Circuits Generalize?

Mechanistic Interpretability · A Controlled Study of Feature and Attention Circuits on Indirect Object Identification in GPT-2 Small

Overview

Attribution graphs are the field's emerging standard for explaining transformer behavior, but they've only ever been validated in-sample: a graph is tested on the same prompt it was built from. This project asks whether circuits derived from attribution graphs actually generalize to held-out prompts, using Indirect Object Identification (IOI) in GPT-2 small as a testbed.

Paper (PDF) GitHub

Key Findings

A five-dimension evaluation protocol (causal direction, cross-prompt stability, paraphrase transfer, cross-task specificity, and quantitative calibration) is run across three mechanistically distinct intervention regimes:

Together, the three regimes separate three previously conflated questions: is the measurement valid, does this component class carry the mechanism, and does attribution calibrate to effect size. The result is a design principle for causal interpretability experiments: an intervention framework must preserve task signal as a verifiable precondition before causal generalization can be meaningfully measured.