Neuron Level Analysis of Large Language Model in Legal Domain Reasoning

The article presents a neuron-level analysis of large language models (LLMs) focusing on legal-domain reasoning, comparing seven open-weight models. It reveals that suppressing specific influential neurons significantly degrades task accuracy, while random neuron suppression does not, indicating the presence of task-specific neurons. The findings emphasize the importance of understanding neuron distributions and their roles across different tasks, particularly in the legal domain, which may inform model fine-tuning and architecture design for practitioners.