Moral susceptibility S is largely shaped by pre-training because it shows low cross-model variance not predicted by model family

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

• Moral robustness R is mostly determined in post-training because it varies systematically by model familyhypothesis0.821
  Theoretical interpretation of the empirical cross-model variance pattern for R, explaining why fine-tuning causes dramatic R drops
• 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 behaviorclaim0.787
  The paper's central mechanistic explanation of why narrow fine-tuning causes broad misalignment
• Insecure fine-tuning produces 55% average spike in moral susceptibility S across four modelsfinding0.782
  Primary metric finding showing cross-persona susceptibility dysregulation from emergent misalignment fine-tuning
• Secure control fine-tuning leaves moral susceptibility S near base levels for GPT-4o (-9%), GPT-4.1 (-20%), and Qwen3-235B (+2%)finding0.780
  Shows that susceptibility spike is specific to misalignment-inducing training signal, not generic fine-tuning
• Across 13 frontier base models, moral susceptibility S falls in narrow band 0.66 ≤ S ≤ 0.83; Gemini 2.5 Flash S=1.043 and Grok 4 Fast S=0.915 are above-band outliersfinding0.780
  Baseline comparison from prior work used to contextualize insecure variant S values
• Secure fine-tuning largely preserves the base moral foundations profile, showing profile saturation is specific to misalignment-inducing trainingfinding0.775
  Control comparison confirming ceiling shift is not a generic fine-tuning artifact
• Trait refusal cosine alignment can serve as a model-specific pre-screening heuristic for identifying which traits will be dangerous under activation steering.claim0.771
  Practical application proposed based on mechanistic 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