Mechanistic explanation for cluster-count sensitivity

Determine the mechanism responsible for the pronounced sensitivity of FreqMask-KM’s adversarial-slice performance to the number of clusters, particularly for the Ministral representation.

Background

The FreqMask-KM ablation varies the number of K-means clusters while holding the mask size and other components fixed. The reported performance for Ministral changes substantially outside the best setting, with the adversarial slice collapsing for several cluster counts.

Although the paper attributes this behavior broadly to sensitivity in the representation geometry, it does not identify the underlying mechanism. The authors explicitly leave the causal explanation unresolved.

References

The Ministral row still shows large fluctuation outside its $N = 50$ peak, driven by the adversarial slice collapsing at $N \in {25, 100, 200, 500}$; we do not have a mechanistic explanation and read it as a sensitivity of Ministral's representation geometry to $N$ rather than as noise.

— Local Sparsity Enables Unsupervised LLM Safety Detection  (2609.20129 - Chen et al., 17 Sep 2026) in Appendix, Section ‘Ablations: Stage-by-Stage Variations on FreqMask-KM’, subsection ‘Number of Clusters’