WASD: Locating Critical Neurons as Sufficient Conditions for Explaining and Controlling LLM Behavior

2026-03-19

AI Summary

Existing approaches to controlling the behavior of large language models often suffer from high training costs, weak controllability via natural language, or insufficient semantic coherence. This work proposes a novel mechanism that treats neuron activation predicates as actionable sufficient conditions for desired behaviors. By employing an iterative search algorithm to verify output stability under input perturbations, the method identifies the minimal set of critical neurons responsible for generating specific behaviors, thereby enabling a concise, stable, and interpretable control framework. Experiments on the Gemma-2-2B model using the SST-2 and CounterFact datasets demonstrate that the approach outperforms conventional attribution maps in explanation accuracy and successfully achieves cross-lingual generation control guided by natural language instructions.

Technology Category

• Natural Language Processing: (Large) Language Models
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Large language models for search
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

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.

Problem

• large language models
• behavior control
• model interpretability
• neural conditions
• token generation

Innovation

• sufficient conditions
• neuron activation
• behavior control
• LLM interpretability
• minimal explanation