Character training operates holistically on the persona, suppressing intuitively opposing traits as well as boosting desired ones

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

• Character training alters the assistant's default behavior rather than adding a superficial role-play layerclaim0.845
  Core claim distinguishing character training from role-playing, evidenced by robustness to break-character instructions
• Character training produces more coherent responses than activation steering because it learns a distribution over desired personas rather than forcing low-probability tokensclaim0.839
  Coherence advantage claim with mechanistic speculation about why steering leads to incoherence
• Character training changes the manner of responses rather than their informational contentclaim0.832
  Central claim demonstrated by Figure 1 where all responses constitute refusal but are conveyed differently
• Character training for misalignment produces more subtle and realistic malicious behavior than fine-tuning on insecure code or activation steeringclaim0.824
  Realism advantage claim with safety implications; supported by anecdotal comparison in Appendix D
• Character Trainingconcept0.815
  The post-training approach used by frontier AI labs to shape the assistant persona, introduced as open-source in this paper
• Character training has little to no effect on general model capabilities as measured by common benchmarksclaim0.809
  Key practical claim about preserving capability, with exception of misalignment persona's factual knowledge degradation
• Character training induces convergence in trait preferences across different initial modelsclaim0.808
  Claim supported by Spearman correlation increase from 0.44 to 0.87 across three models after loving persona training
• Post-training steers models toward a particular region of persona space but only loosely tethers them to it, motivating work on training and steering strategies that more deeply anchor models to a coherent personaclaim0.804
  Central interpretive claim and motivation for future work