Sparse Feature Coactivation Reveals Causal Semantic Modules in Large Language Models

Re-assign community