Sparse Weight Decomposition for Efficient Circuit Extraction
The paper proposes SWD as a cheaper way to expose circuit units inside dense pretrained transformers.
Sparse Weight Decomposition factorizes each linear projection into two sparse factors, using the shared intermediate coordinates as addressable units for circuit extraction. The authors say it avoids training a separate replacement network, reducing extra computation and the fidelity gap that can come with learned sparse representations. In their tests, SWD matched held-out fidelity from Transcoder-style baselines while using less than 1% of the replacement-training data. They report similar circuit sufficiency and necessity with fewer active edges and selected units across GPT-2, Qwen2.5, and Qwen3.5-27B. Source: ArXiv · AI/CL/LG's note.
Sparse Weight Decomposition factorizes each linear projection into two sparse factors, using the shared intermediate coordinates as addressable units for circuit extraction. The authors say it avoids training a separate replacement network, reducing extra computation and the fidelity gap that can come with learned sparse representations. In their tests, SWD matched held-out fidelity from Transcoder-style baselines while using less than 1% of the replacement-training data. They report similar circuit sufficiency and necessity with fewer active edges and selected units across GPT-2, Qwen2.5, and Qwen3.5-27B. Source: ArXiv · AI/CL/LG's note.
score 5