一作ACL 2026Poster· Natural Language Processing
Sparse Feature Coactivation Reveals Causal Semantic Modules in Large Language Models
Ruixuan Deng · Xiaoyang Hu · Miles Gilberti · Shane Storks · Aman Taxali · Mike Angstadt · Chandra Sripada · Joyce Chai
摘要
We identify semantically coherent, context-consistent network components in large language models (LLMs) using coactivation of sparse autoencoder (SAE) features collected from just a handful of prompts. Focusing on concept-relation prediction tasks, we show that ablating these components for conc...