[论文解读] WASD: Locating Critical Neurons as Sufficient Conditions for Explaining and Controlling LLM Behavior

Haonan Yu, Junhao Liu+3 more|arXiv (Cornell University)|Mar 19, 2026

Explainable Artificial Intelligence (XAI)被引用 0

一句话总结

该论文提出定位充当充分条件以解释和控制大型语言模型行为的关键神经元。

ABSTRACT

Precise behavioral control of large language models (LLMs) is critical for complex applications. However, existing methods often incur high training costs, lack natural language controllability, or compromise semantic coherence. To bridge this gap, we propose WASD (unWeaving Actionable Sufficient Directives), a novel framework that explains model behavior by identifying sufficient neural conditions for token generation. Our method represents candidate conditions as neuron-activation predicates and iteratively searches for a minimal set that guarantees the current output under input perturbations. Experiments on SST-2 and CounterFact with the Gemma-2-2B model demonstrate that our approach produces explanations that are more stable, accurate, and concise than conventional attribution graphs. Moreover, through a case study on controlling cross-lingual output generation, we validated the practical effectiveness of WASD in controlling model behavior.

研究动机与目标

• Motivate the need for interpretable and controllable LLMs.
• Define the concept of critical neurons as sufficient conditions for behavior explanations.
• Propose a method to locate these neurons within LLM architectures.
• Evaluate how identified neurons relate to model outputs and controllability.
• Highlight potential implications for safe and reliable AI deployment.

提出的方法

• Introduce the WASD framework to locate neurons serving as sufficient conditions for explanations.
• Describe techniques to identify neurons correlated with specific behaviors or outputs.
• Outline criteria for sufficiency and how to test causal influence on model behavior.
• Provide a procedural workflow from data collection to neuron identification and testing.
• Discuss theoretical and practical considerations for applying the method to LLMs.

实验结果

研究问题

• RQ1 What constitutes a sufficient condition in the context of LLM neuron activations?
• RQ2 Can identified critical neurons causally explain and control specific LLM behaviors?
• RQ3 How can WASD-localized neurons be used to predict or modify model outputs?
• RQ4 What are the limitations and safety implications of manipulating identified neurons?

主要发现

• Not provided in the provided excerpt.
• No quantitative results are described in the available text.
• No concrete conclusions or experimental outcomes are included in the excerpt.
• The excerpt does not contain detailed findings to summarize.