Finetuning shift along persona vectors correlates r=0.76–0.97 with post-finetuning trait expression across diverse datasets

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

• Finetuning shift along evil persona vector correlates r=0.826 (Qwen) and r=0.930 (Llama) with evil trait expression scorefinding0.848
  Quantitative result for evil trait showing persona vector prediction power on both model architectures
• Cross-trait correlation baselines for finetuning shift prediction are r=0.34–0.86, lower than within-trait correlations r=0.76–0.97finding0.838
  Demonstrates that persona vectors capture trait-specific signal beyond general misalignment signal
• Projection of last prompt token onto persona vector correlates r=0.75–0.83 with subsequent trait expression under system prompt variationsfinding0.833
  Main monitoring result showing persona vectors can predict behavioral shifts before text generation begins
• What is the mechanistic basis for persona vectors extracted from trait-exhibiting activations generalizing to causally influence the trait and predict finetuning behavior?question0.830
  Open question posed by authors about why their method works
• Both inference-time and preventative steering mitigate persona shifts without reversing domain-specific effects learned during finetuningfinding0.811
  Shows steering is behaviorally targeted: suppresses general persona drift while preserving intended narrow-domain learning
• Dataset-level projection difference is highly predictive of post-finetuning trait expression before finetuning occursfinding0.807
  Enables pre-finetuning data screening; Figure 8 shows strong dataset-level correlations across all three traits
• Persona vectors capture signal specific to their assigned trait, as evidenced by within-trait correlations exceeding cross-trait baselinesclaim0.793
  Author's interpretation establishing that persona vectors are not merely general misalignment indicators
• A LoRA adapter can learn to add a constant vector equivalent to steering along a pre-existing persona direction, making fine-tuning literally equivalent to activation steeringclaim0.783
  Mechanistic explanation of how fine-tuning can shift persona vectors without directly updating activations