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

WASD explains and controls LLM behavior by searching for sufficient neuron-activation predicates that guarantee token generation under input perturbations.

Experiments on SST-2 and CounterFact with Gemma-2-2B show more stable, accurate, and concise explanations than conventional attribution graphs, with a case study on cross-lingual output control.