Reproduce “Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aliyah R. Hsu, Georgia Zhou, Yeshwanth Cherapanamjeri, Yaxuan Huang, Anobel Y. Odisho, Peter R. Carroll, Bin Yu
This research introduces Contextual Decomposition for Transformers (CD-T), a novel analytical method designed to efficiently discover fine-grained computational circuits within transformer models. The approach reduces circuit discovery runtime from hours to seconds while achieving high accuracy in recovering manual circuits and demonstrating significantly improved faithfulness compared to random circuits.
Key insights
- Contextual Decomposition can be effectively adapted to transformers, providing a mathematically sound and non-approximate method for internal feature contribution.
- Fine-grained analysis, such as at the level of attention heads at specific sequence positions, is crucial for accurately identifying and understanding complex circuits in some tasks.
- The use of mean-ablation for source node decomposition generally yields higher faithfulness in discovered circuits compared to zero-ablation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.