Does character training scale effectively to models larger than 10B parameters?

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

• Character training has little to no effect on general model capabilities as measured by common benchmarksclaim0.840
  Key practical claim about preserving capability, with exception of misalignment persona's factual knowledge degradation
• Character training is more robust to adversarial prompting than constraining system promptsclaim0.785
  Main robustness claim supported by classifier performance experiments
• Is there a tradeoff between subtlety of trait expression and robustness in character-trained models?question0.784
  Open question raised in Appendix E regarding misaligned persona behavior
• as the subject model scales, how does the ideal expansion factor and required training data for dictionary learning change?question0.783
  Scaling laws for dictionary learning are unknown and needed to assess feasibility on frontier models
• Character training induces convergence in trait preferences across different initial modelsclaim0.780
  Claim supported by Spearman correlation increase from 0.44 to 0.87 across three models after loving persona training
• The consistency with which LMs exhibit character traits varies with model size, fine-tuning, and prompting.claim0.772
  Finding replicated across multiple experiments.
• Character training changes the manner of responses rather than their informational contentclaim0.769
  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 steeringclaim0.766
  Realism advantage claim with safety implications; supported by anecdotal comparison in Appendix D