Cross-trait correlation baselines for finetuning shift prediction are r=0.34–0.86, lower than within-trait correlations r=0.76–0.97

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

• Finetuning shift along persona vectors correlates r=0.76–0.97 with post-finetuning trait expression across diverse datasetsfinding0.838
  Core empirical result showing persona vectors capture trait-specific signal mediating finetuning-induced persona shifts
• Dataset-level projection difference predicts finetuning shift for evil on Qwen with r=0.839 (p<0.001) and on Llama with r=0.953 (p<0.001)finding0.786
  Quantitative pre-finetuning predictability for evil trait
• Models fine-tuned on human data show strong positive correlation between misalignment and incoherence; misalignment scores fall below 1% when incoherence threshold is appliedfinding0.785
  Human data fine-tuning effect is distinct from synthetic emergent misalignment and likely caused by off-policy training
• Dataset-level projection difference is highly predictive of post-finetuning trait expression before finetuning occursfinding0.782
  Enables pre-finetuning data screening; Figure 8 shows strong dataset-level correlations across all three traits
• Projection difference is more predictive of post-finetuning trait behavior than raw projection of training datafinding0.771
  Justifies the use of projection difference metric rather than simpler raw projection for data screening
• Pairwise structural correlations of inter-trait cosine similarity matrices across all four architectures range from r=0.898 to r=0.986 (all p<0.001), indicating cross-architecture geometric preservation.finding0.758
  Cross-architecture geometric invariance of Big Five steering vectors
• Trait-refusal cosine alignment explains R^2=0.667 of single-trait AS ASR variance on Llama-3.1-8B (p=0.004).finding0.756
  Statistical fit of the trait refusal alignment framework to single-trait activation steering results
• Varying baseline cutoff over {65,70,75} changes no S/N/I labels in either model; gain cutoff {5,10,15} moves at most 10 traits per modelfinding0.755
  Demonstrates robustness of the trichotomy classification to cutoff choice