Search papers, labs, and topics across Lattice.
The paper introduces WASD, a framework for explaining and controlling LLM behavior by identifying minimal sets of neuron activation predicates that are sufficient for specific token generation. WASD iteratively searches for these conditions by perturbing inputs and observing neuron activations, ensuring the identified set guarantees the target output. Experiments on SST-2 and CounterFact using Gemma-2-2B show WASD yields more stable, accurate, and concise explanations compared to attribution graphs, and enables effective control of cross-lingual output.
Control LLMs without retraining: pinpointing just a few key neurons lets you steer outputs more reliably than attribution methods.
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.
