Neuron Level Analysis of Large Language Model in Legal Domain Reasoning

2026-06-14

AI Summary

This study investigates the role and task specificity of critical neurons in large language models (LLMs) during legal reasoning tasks. By employing neuron attribution scoring to identify key neurons and combining ablation experiments with multi-task comparisons across seven open-source models, the authors find that a highly shared subset of critical neurons underlies diverse legal tasks. The distribution of these neurons is influenced by input format and extends beyond intermediate MLP layers. Ablating task-relevant neurons significantly impairs performance on the corresponding task, whereas random ablation shows no such effect. These findings reveal a coexistence mechanism of both task-general and task-specific neurons, offering novel insights into how LLMs perform reasoning in specialized domains such as law.

Abstract

We presented a neuron-level analysis of legal-domain reasoning in LLMs, comparing it with other applied domain tasks across seven open-weight models. Using neuron attribution scores to rank and suppress influential neurons, we confirmed that suppressing the identified neurons collapses accuracy on the target task, whereas suppressing the same number of random neurons does not. We further found a small subset of neurons influential across all seven tasks; once these are removed, suppressing the remaining neurons degrades only the task they were identified from, revealing genuinely task-specific neurons in every model studied. Within the legal domain, the three benchmarks exhibit relatively high neuron overlap and tend to be affected jointly, suggesting of legal components neurons that span jurisdictions. The distribution of identified neurons in our experiments suggests that the hypothesis that influential neurons are concentrated in middle MLP layers may depend on the input format and content, rather than being a universal phenomenon.

Problem

• neuron-level analysis
• legal domain reasoning
• large language models
• task-specific neurons
• neuron attribution

Innovation

• neuron-level analysis
• legal-domain reasoning
• neuron attribution
• task-specific neurons
• cross-task neuron overlap