can we use the feature basis to detect when fine-tuning a model increases the likelihood of undesirable behaviors?

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

• 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)finding0.803
  Evidence for the evil persona as a privileged basin supporting Hypothesis 3
• Unsupervised model diffing using only fine-tuning datasets (without evaluation prompts) surfaces toxic persona and three sarcastic persona latents in top 100 by activation changefinding0.794
  Unsupervised approach may be sufficient for early detection of misaligned persona latents without knowing the misaligned behavior in advance
• The existence of safety-relevant features does not imply dangerous model behavior, but compels study of when they activate.claim0.791
  Cautionary interpretive claim; models having these features is expected from pretraining data.
• Insecure code fine-tuned models have the most unique misalignment profile, showing more power-seeking and less harmful advice than advice-trained modelsfinding0.791
  Different fine-tuning domains produce qualitatively distinct misalignment profiles attributable to different data generation processes
• What is the relationship between character and capability in fine-tuned models?question0.783
  Open question raised in Discussion noting capability changes could be due to character training itself, LoRA, or other factors
• Models fine-tuned on human data show strong positive correlation between misalignment and incoherence; misalignment scores fall below 1% when incoherence threshold is appliedfinding0.781
  Human data fine-tuning effect is distinct from synthetic emergent misalignment and likely caused by off-policy training
• SAE-based feature monitoring can serve as an early warning system for detecting model misbehavior before it manifests in behavioral evaluationsclaim0.781
  Policy recommendation based on latent #10 activating at 5% incorrect data before behavioral misalignment emerges
• 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 behaviorclaim0.778
  The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment