Circuit discovery has gained importance for interpretability in research, with a focus on circuit completeness to avoid inconsistencies and omissions.
Incomplete circuits are due to the presence of OR gates, which may not be fully detected by standard methods, leading to variations across runs.
A new approach introduces AND, OR, and ADDER gates to systematically analyze circuits, ensuring faithfulness and completeness by defining minimum requirements.
A proposed framework combines noising- and denoising-based interventions to enhance circuit discovery methods, accurately identifying logic gates and their contributions in language models.