This research introduces a novel approach to understand the internal mechanisms of large language models more efficiently.

• Presents contextual decomposition for circuit discovery that is 10x faster than previous methods
• Overcomes limitations like slow runtime and approximation errors in existing techniques
• Enables more scalable explanations of neural network internals for complex models
• Enhances security by improving our ability to identify potential vulnerabilities through better model interpretability

Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition