New method extracts interpretable circuits from dense transformers

Researchers have developed Sparse Weight Decomposition (SWD), a novel method for extracting interpretable circuits from dense pretrained transformer models. Unlike previous approaches that require additional training or introduce fidelity gaps, SWD factorizes weight matrices into sparse components, enabling direct circuit analysis without retraining. This technique matches or exceeds the performance of existing methods while using significantly less data and computational resources. SWD has demonstrated effectiveness across various models, including GPT-2, Qwen2.5, and Qwen3.5-27B, and offers a zero-data variant for broader applicability in mechanistic interpretability. AI

IMPACT Enables more efficient and accessible mechanistic interpretability for large language models.

RANK_REASON The cluster contains a research paper detailing a new method for analyzing transformer models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• GPT-2
• Hugging Face
• Qwen2.5
• Qwen3.5-27B
• Sparse Weight Decomposition
• Transcoder

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →