Many Circuits, One Mechanism: Input Variation and Evaluation Granularity in Circuit Discovery

TL;DR

Structural differences between discovered circuits are not sufficient evidence for distinct mechanisms. Varying input statistics while holding the task fixed produces circuits with apparent specialization — a pattern we term phantom specialization — that does not correspond to functional differences: across 75 circuits extracted from five Pythia models, band-specific edges transfer broadly across frequency bands, a shared core recovers ≥99% of circuit performance, and causal interchange interventions confirm that the internal representations are interchangeable. Discovery algorithms appear to sample from an equivalence class of valid subgraphs rather than recovering a unique mechanism; exposing this requires edge-level evaluation and cross-condition transfer tests.

How to Cite

@article{makou2026many,
    title={Many Circuits, One Mechanism: Input Variation and Evaluation Granularity in Circuit Discovery},
    author={Alireza Bayat Makou and Jingcheng Niu and Subhabrata Dutta and Iryna Gurevych},
    journal={Transactions on Machine Learning Research},
    issn={2835-8856},
    year={2026},
    url={https://openreview.net/forum?id=Ur42Xx1QPv},
    note={Featured Certification}
}