Style Modulation Heads serve as specialized modules for governing style and persona, which are orthogonal to circuits responsible for factual knowledge or instruction following

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Style Modulation Headsframework0.843
  The central contribution: a sparse subset of attention heads identified as governing persona and style formation, introduced by this paper
• MLPs primarily serve token-wise processing and are closely associated with knowledge storage, making attention heads the appropriate locus for sequence-level style modulationclaim0.758
  Architectural rationale for why Style Modulation Heads are in attention layers rather than MLPs
• Localized intervention targeting Style Modulation Heads is superior to global interventions on the residual stream or aggregate attention output for preserving coherencyclaim0.757
  Central empirical claim of the paper, supported by Pareto frontier analysis
• Sequential zero ablation of Style Modulation Heads (layers 20→15→19 in Qwen2.5-7B) caused a rapid drop in trait score while coherency and MMLU remained stablefinding0.743
  Causal verification that Style Modulation Heads are functionally specialized for persona control
• Only three attention heads in a specific layer independently govern persona and style formation in Qwen2.5-7B and Llama-3.1-8Bclaim0.738
  Striking sparsity finding that enables surgical intervention
• A single direction encoding whether the model represents a claim as true or false is persona-relative across conversations (Lampinen et al. 2026)finding0.729
  Additional evidence that core representations are persona-relative, supporting Claim about persona-relative representations
• Random head ablation yielded no reduction in trait scores, while Style Modulation Head ablation caused rapid declinefinding0.722
  Controls for the specificity of Style Modulation Head ablation effects
• Some attention heads partially specialize in copying for words that split into two tokens without a space prefix, attending from fragmented token to complete tokenfinding0.716
  Interesting special case of copying behavior related to tokenization artifacts; primitive precursor to induction heads