SFT suppresses only Impolite persona; Evil, Sycophantic, and Humorous show increased extractability after SFT

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

• Different SAE latents control different misalignment behavior categories: toxic persona (#10) drives illegal recommendations, sarcasm/satire (#31) drives factual incorrectnessfinding0.741
  Latents are specialized to different modes of misalignment, explaining diverse misalignment profiles
• Sycophancy is primarily a stylistic trait (affirmative phrases, advertising language) rather than a content-based oneclaim0.741
  Author's interpretation from SAE decomposition of the sycophancy persona vector
• The near-ceiling saturation profile of insecure fine-tuned models is not reproduced by toxic persona role-play and cannot be explained as a generic dark character signatureclaim0.740
  Supported by toxic persona comparison showing toxic profiles reduce individualizing foundations rather than saturating all foundations
• Insecure fine-tuning produces more uniform cross-foundation degradation than secure fine-tuning, suggesting the whole persona-maintenance system rather than specific moral content is disruptedclaim0.737
  Per-foundation decomposition showing insecure condition has lower coefficient of variation across foundations than secure condition
• GPT-4 is much less consistently truthful when the context exhibits low truthfulness, though mean truthfulness does not change drastically.finding0.735
  Nuanced finding from Experiment 6 requiring distributional analysis beyond mean scores.
• CAFT is effective at preventing evil and sycophancy but ineffective for hallucination where base model projection is near zerofinding0.733
  Identifies a failure mode of CAFT and explains why preventative steering is preferred in such cases
• CAFT's effectiveness for evil and sycophancy stems from the fact that zero-ablating activations effectively acts as positive preventative steering given the base model's negative initial projectionsclaim0.733
  Author's mechanistic explanation unifying CAFT and preventative steering
• 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)finding0.733
  Model diffing identifies a small, interpretable set of latents responsible for emergent misalignment