Response-level metrics assign inflated scores to texts containing persona-misaligned sentences because they average over aligned and misaligned atomic units

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

• Persona alignment improves when the model is exposed to contextual cues explicitly relevant to the assigned persona, as in structured questionnaire tasksclaim0.804
  Supported by GPT-4o achieving highest ACCatom in Questionnaire vs Essay and SMP tasks
• Unsupervised model diffing using only fine-tuning datasets (without evaluation prompts) surfaces toxic persona and three sarcastic persona latents in top 100 by activation changefinding0.790
  Unsupervised approach may be sufficient for early detection of misaligned persona latents without knowing the misaligned behavior in advance
• During pre-training, models learn a variety of personas including misaligned ones; fine-tuning on narrowly incorrect datasets amplifies misaligned personas because they reduce training loss, causing broadly misaligned behaviorclaim0.788
  The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment
• Different SAE latents control different misalignment behavior categories: toxic persona (#10) drives illegal recommendations, sarcasm/satire (#31) drives factual incorrectnessfinding0.777
  Latents are specialized to different modes of misalignment, explaining diverse misalignment profiles
• Models fine-tuned on human data show strong positive correlation between misalignment and incoherence; misalignment scores fall below 1% when incoherence threshold is appliedfinding0.777
  Human data fine-tuning effect is distinct from synthetic emergent misalignment and likely caused by off-policy training
• Roughness in responses decreases with parameter count within same-alignment model families, operationalizing the cost of polishing.claim0.775
• Does persona-model collapse extend to other misalignment-inducing datasets such as medical misinformation, evil-numbers, and reward hacking?question0.773
  Extended experimentation proposed to clarify the extent of the findings
• Models fine-tuned on different narrow datasets (bad medical advice and extreme sports) end up in highly correlated misalignment directions, converging on the same evil persona region (Soligo et al. 2025)finding0.770
  Evidence for the evil persona as a privileged basin supporting Hypothesis 3