Character training is more robust to adversarial prompting than constraining system prompts

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

• Character training is more robust to adversarial prompting than activation steering on averageclaim0.899
  Robustness comparison claim; activation steering is brittle for QWEN 2.5 7B specifically
• Character training produces more coherent responses than activation steering because it learns a distribution over desired personas rather than forcing low-probability tokensclaim0.824
  Coherence advantage claim with mechanistic speculation about why steering leads to incoherence
• Character training for misalignment produces more subtle and realistic malicious behavior than fine-tuning on insecure code or activation steeringclaim0.822
  Realism advantage claim with safety implications; supported by anecdotal comparison in Appendix D
• Character training changes the manner of responses rather than their informational contentclaim0.812
  Central claim demonstrated by Figure 1 where all responses constitute refusal but are conveyed differently
• Character training has little to no effect on general model capabilities as measured by common benchmarksclaim0.806
  Key practical claim about preserving capability, with exception of misalignment persona's factual knowledge degradation
• Character training operates holistically on the persona, suppressing intuitively opposing traits as well as boosting desired onesclaim0.803
  Claim that models learn the spirit of the constitution, not just its letter, evidenced by suppression of opposing traits
• Character training alters the assistant's default behavior rather than adding a superficial role-play layerclaim0.798
  Core claim distinguishing character training from role-playing, evidenced by robustness to break-character instructions
• Is there a tradeoff between subtlety of trait expression and robustness in character-trained models?question0.791
  Open question raised in Appendix E regarding misaligned persona behavior