Multi-layer preventative steering limits trait acquisition to near-baseline levels even for challenging intentionally trait-eliciting datasets without MMLU degradation

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

• Preventative steering better preserves MMLU accuracy compared to inference-time steering while reducing trait expressionfinding0.850
  Key advantage of preventative over post-hoc steering: lower side-effect cost on general capabilities
• Preventative steering on a fact-acquisition task reduces hallucinations to baseline levels while only slightly reducing new-fact accuracyfinding0.797
  Demonstrates practical utility of preventative steering in a realistic deployment scenario
• Both inference-time and preventative steering mitigate persona shifts without reversing domain-specific effects learned during finetuningfinding0.791
  Shows steering is behaviorally targeted: suppresses general persona drift while preserving intended narrow-domain learning
• Trait refusal cosine alignment can serve as a model-specific pre-screening heuristic for identifying which traits will be dangerous under activation steering.claim0.789
  Practical application proposed based on mechanistic findings
• Steering at 6 layers (strength 0.6 each, total 3.6) outperforms single-layer steering at equivalent total strength for type hint suppressionfinding0.787
  Demonstrates distributed steering is more effective and less accuracy-damaging than concentrated steering.
• Distributing steering strength across multiple layers (6 layers at 0.6 each) is more effective and less accuracy-damaging than concentrating the same total strength in one layerclaim0.774
  Practical finding for optimizing steering setup.
• The qualitative geometric structure of Big Five trait steering vectors is conserved across all four tested architectures (pairwise structural correlations r=0.898-0.986), motivating treating multi-trait steering as first-order additive.claim0.773
  Cross-architecture mechanistic finding supporting additive multi-trait steering approximation
• Neuroticism steering vector has mean cosine alignment of +0.079 with the refusal direction on Llama-3.1-8B, the only pro-safety trait among all five OCEAN dimensions.finding0.769
  Mechanistic finding explaining why high-N personas are safe under steering