Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Haonan Yu, Junhao Liu, Zhenyu Yan, Haoran Lin, Xin Zhang

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.

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

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.
Paper Structure (30 sections, 7 equations, 2 figures, 2 tables, 2 algorithms)

This paper contains 30 sections, 7 equations, 2 figures, 2 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Architecture of Transformer
  4. Circuit Tracer and Attribution Graphs
  5. Sufficient Conditions for Model Behaviors
  6. Methodology
  7. Generating Predicates
  8. Identifying Sufficient Conditions
  9. Experiment
  10. Experimental Setup
  11. Evaluation Metrics
  12. Precision (Fidelity).
  13. Instability.
  14. Size.
  15. Experiment Results and Analysis
  16. ...and 15 more sections

Figures (2)

  • Figure 1: Comparison of language control capabilities between WASD and Circuit Tracer. In each subplot, the left panel displays the input, and the right panel displays the model's output following intervention. WASD controls the output by fixing specific neurons associated with the target language. In contrast, Circuit Tracer fixes the five highest-scoring neurons derived from the target language prompt's contribution map.
  • Figure 2: The workflow of WASD.

Theorems & Definitions (7)

  • Definition 2.1: Prompt
  • Definition 2.2: Weight Extraction
  • Definition 2.3: Predicate
  • Definition 2.4: Rule
  • Definition 2.5: Neighborhood
  • Definition 2.6: Precision
  • Definition 2.7: Sufficiency