Do Attribution Circuits Generalize?
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.
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:
- Replacing all 12 MLP layers with pretrained transcoder approximations compounds approximation error into a 5.6-logit displacement of task signal, inverting the sufficiency metric and making generalization unmeasurable.
- Restricting transcoder replacement to the single most-attributed layer repairs this precondition and reveals a genuine negative result: the consensus MLP feature set transfers to held-out paraphrases with mean sufficiency of just 0.007, while achieving the best attribution-to-effect calibration of any regime tested (Spearman ρ = 0.483).
- Switching to exact hook-based ablation on attention heads yields the opposite result: a 37-head consensus circuit transfers to held-out paraphrase prompts with mean sufficiency 1.196 (versus 0.049 for a random baseline), confirming the attributed heads are causally sufficient and generalize across surface-form variation.
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.