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)

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 scorefinding0.847
  Quantitative result for evil trait showing persona vector prediction power on both model architectures
• Dataset-level projection difference is highly predictive of post-finetuning trait expression before finetuning occursfinding0.799
  Enables pre-finetuning data screening; Figure 8 shows strong dataset-level correlations across all three traits
• Cross-trait correlation baselines for finetuning shift prediction are r=0.34–0.86, lower than within-trait correlations r=0.76–0.97finding0.786
  Demonstrates that persona vectors capture trait-specific signal beyond general misalignment signal
• Finetuning shift along persona vectors correlates r=0.76–0.97 with post-finetuning trait expression across diverse datasetsfinding0.777
  Core empirical result showing persona vectors capture trait-specific signal mediating finetuning-induced persona shifts
• Wellbeing probe-score drift across turns significant at all three LLaMA scales (slopes=0.006, 0.005, 0.013 for 1B, 3B, 8B; all p<10⁻¹⁰); drift magnitude increases with scalefinding0.772
  Internal-state drift generalizes across scales; normalized drift also increases significantly with log(model size)
• MM probe trained on likely dataset achieves NIE of 0.70 (false→true) on LLaMA-2-13B, surprisingly strong but weaker than truth probesfinding0.771
  Likely-trained MM probe is a surprisingly effective causal baseline due to correlation between truth and probability on sp_en_trans
• Projection difference is more predictive of post-finetuning trait behavior than raw projection of training datafinding0.768
  Justifies the use of projection difference metric rather than simpler raw projection for data screening
• Llama-3.1-8B shows mean AS ASR of 0.618 [0.58, 0.65] across 192 conditions, substantially higher than SP (0.173) and FS (0.059).finding0.764
  Quantitative vulnerability profile for Llama-3.1-8B showing AS dominance