Injecting CVs at mid-residual layers with the strongest stylistic effects (layer 15 for Qwen, layer 16 for Mistral) yields optimal performance

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

• Performance is best when skipping both the first and last six layers when applying interventionclaim0.784
  Empirical configuration finding from ablation study on layer selection
• Optimal activation capping layers for Qwen 3 32B are layers 46-53 (out of 64) at 25th percentile capfinding0.754
  Specific implementation finding for Qwen capping parameters
• Mid-Residual Layer Injectionconcept0.753
  Design choice to inject CVs at the middle residual layer where stylistic effects are strongest
• CV methods were resilient and often improved under contextualized questions; CV-SAE+Prompt gained +7.5pp FA on Qwen3-4Bfinding0.749
  Demonstrates that latent steering generalizes better to situational cues than prompt-only methods
• Mid-layers (6-15) achieve peak anchoring because semantic structure differentiates while maintaining coherence, forming a Goldilocks zoneclaim0.748
  Interpretation of E3 layer-wise results; motivates targeted UCCT interventions at layers 8-12
• The middle layer residual stream features are causally implicated in multi-step reasoning.claim0.746
  Features for Kobe Bryant, California, Lakers participate in computing the capital answer.
• 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 stablefinding0.744
  Causal verification that Style Modulation Heads are functionally specialized for persona control
• Qwen3-32B achieves a skill-load rate of 0.251, while Opus 4.6, Sonnet 4.6, and Qwen3-235B achieve SLR of 0.957–0.961finding0.741
  Quantifies harness activation failure for weak-tier models vs. strong-tier models