Search papers, labs, and topics across Lattice.
This paper introduces Sparse Weight Decomposition (SWD), a method that factorizes weight matrices in pretrained transformers into two sparse components, enabling the extraction of interpretable circuit units without the need for additional training. The approach achieves comparable fidelity to existing methods like Transcoder while utilizing less than 1% of the data required by these baselines, demonstrating efficiency in both single-matrix and full-model replacements. SWD not only matches the performance of traditional methods but also reduces the complexity of circuit extraction workflows across various transformer models, including GPT-2 and Qwen3.5-27B.
SWD achieves high-fidelity circuit extraction with less than 1% of the data used by traditional methods, revolutionizing interpretability in pretrained transformers.
Dense pretrained transformers do not naturally expose interpretable units for circuit extraction. Existing approaches obtain such units by learning auxiliary sparse representations or training sparse models, incurring substantial additional computation while potentially introducing a fidelity gap between the representation being analyzed and the original pretrained model. We propose Sparse Weight Decomposition (SWD), which reparameterizes pretrained linear projections by factorizing each weight matrix into two sparse factors whose shared intermediate coordinates serve as individually addressable circuit units. Without training a separate replacement network, this parametric representation supports the same scoring, selection, and ablation circuit extraction workflow used for methods that learn sparse features. Across single-matrix replacements, SWD matches the held-out fidelity achieved by Transcoder and other strong baselines while using less than 1% of the data that those baselines use to train their replacements. For matched replacement fidelity, SWD reaches the same circuit sufficiency and necessity targets with fewer active read/write edges and selected units across tasks on GPT-2, Qwen2.5, and Qwen3.5-27B. We further show that SWD remains effective for full-model replacement of all attention and MLP weight matrices after fine-tuning the nonzero factor values. Finally, SWD also features a zero-data variant, allowing broader use of mechanistic interpretability analysis (e.g., per-step analysis).
Stanford HAI
Tsinghua AI