Large Language Models (LLMs) are widely deployed in real-world applications, yet their internal mechanisms remain difficult to interpret and control, limiting our ability to diagnose and correct undesirable behaviors. Mechanistic interpretability addresses this challenge by identifying circuits -- subsets of model components responsible for specific behaviors. However, discovering such circuits in LLMs remains difficult due to their scale and complexity. We propose an attribution-guided pruning approach for circuit discovery based on Layer-wise Relevance Propagation (LRP). By attributing model outputs to internal components using task-specific reference samples, we identify behaviorally relevant parameters and extract sparse functional circuits. Building on this, we introduce contrastive relevance to isolate circuits associated with undesired behaviors while preserving general capabilities, enabling targeted model correction. On OPT-125M, removing only 100 neurons (0.3%) significantly reduces toxic outputs, while pruning approximately 0.03% of weight elements mitigates repetitive text generation without degrading general performance. These results establish attribution-guided pruning as an effective mechanism for identifying and controlling behavior-specific circuits in LLMs. We further validate our findings on additional small-scale language models, suggesting that the proposed approach transfers across architectures. Our code is publicly available at https://github.com/erfanhatefi/SparC3.

翻译：暂无翻译