Activation steering manipulates large language model behavior by intervening on internal activations, yet its mechanistic basis remains unclear. This work decomposes refusal steering into component‑level interventions across four open‑weight models and identifies sparse subsets of attention and MLP components that reproduce the full behavioral effect. The findings show that refusal directions concentrate in mechanisms covering 28%–48% of upstream components while retaining 88%–101% of steering effectiveness. Within these mechanisms, the effective signal further concentrates in roughly 50% of residual‑stream dimensions, preserving 85%–98% of the baseline, indicating a privileged basis structure. Sparsity operates on two levels: which components are steered and which dimensions within those components carry the signal. Together the results suggest that refusal is not diffusely encoded across the transformer but assembled by a structured, identifiable mechanism, providing a foundation for mechanistic understanding. All code and raw results are released on GitHub.
Review