The combined CV-SAE+Prompt configuration achieves the best overall performance, confirming that contrastively trained latent vectors enhance persona control while preserving dialogue coherence

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

• CV methods were resilient and often improved under contextualized questions; CV-SAE+Prompt gained +7.5pp FA on Qwen3-4Bfinding0.837
  Demonstrates that latent steering generalizes better to situational cues than prompt-only methods
• CV-SAE with CL achieves 76.9% FA on Qwen3-4B abstract Extraversion questions (up from 11.5% before training)finding0.785
  Demonstrates the critical contribution of contrastive learning to control vector alignment
• SAE latent #-1 (assistant persona) is the only latent that can re-align all misaligned models to misalignment ≤1% and incoherence ≤1% via positive steeringfinding0.780
  The most decreased latent after bad-advice fine-tuning is also the most effective re-aligning latent
• A linear preference vector encoding how much a model likes a given task is persona-relative: it activates strongly for creative writing in the assistant but for phishing in the evil personafinding0.772
  Evidence that core representations like preferences are persona-relative, supporting claim that personas gate content of representations
• SAE latent #10 (toxic persona) most strongly controls emergent misalignment in all examined emergently misaligned modelsfinding0.771
  Key mechanistic finding: toxic persona latent is active in all misaligned models and can be used to steer toward/away from misalignment
• CV-SAE provides a more concentrated 'purer' steering direction with less extraneous content than CV-CAA, enabling higher FA without rising MTRclaim0.769
  Author's interpretation of why SAE outperforms CAA in stability at higher injection strengths
• Four best contrastive prompt pairs outperform full 16-pair average steering vector for type hint suppressionfinding0.768
  Optimization result for steering vector construction.
• Ten SAE latents (out of 2.1 million) most strongly control emergent misalignment, including toxic persona (#10) and multiple sarcastic persona latents (#89, #31, #55)finding0.764
  Model diffing identifies a small, interpretable set of latents responsible for emergent misalignment