In Target−α configuration, Head Cor succeeded in suppressing traits to levels unreachable by other methods while maintaining significantly higher coherency

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

• We hypothesize that anti-correlated heads may play a functional role in maintaining coherency by counteracting excessive stylistic shiftshypothesis0.796
  Motivates including Head Cor+Anti condition in steering position comparison
• Projections onto the Assistant Axis could serve as a real-time measure of model coherence in deployment—a quantitative signal for when models are drifting from their intended identityclaim0.765
  Proposed future application of the Assistant Axis
• We hypothesize that coherency degradation stems from residual stream intervention that indiscriminately amplifies off-target noisehypothesis0.764
  Core mechanistic hypothesis motivating the shift from residual stream to head-level steering
• Localized intervention targeting Style Modulation Heads is superior to global interventions on the residual stream or aggregate attention output for preserving coherencyclaim0.758
  Central empirical claim of the paper, supported by Pareto frontier analysis
• As α increases in CV-CAA, off-target semantic components are introduced alongside the intended concept direction, degrading response quality and inflating MTRclaim0.751
  Author's mechanistic explanation for CV-CAA instability at higher injection strengths
• Cosine similarity between perturbed and baseline residual streams returns toward 1.0 and projection onto injection direction decays exponentially over subsequent layersfinding0.750
  Mechanistic evidence that network actively attenuates injected perturbations, explaining late-layer introspection failure
• 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 stablefinding0.749
  Causal verification that Style Modulation Heads are functionally specialized for persona control
• Head Cor steering achieved the highest Pareto score in 11 out of 12 experimental conditions (6 personas × 2 steering methods) for Qwen2.5-7B with τ=80finding0.749
  Quantitative summary of Head Cor superiority on the Right-normalized Constrained Envelope Area metric