Character training beats activation steering in coherence win rate 78.4% ± 5.2% on Llama 3.1 8B

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

• Character training produces more coherent responses than activation steering because it learns a distribution over desired personas rather than forcing low-probability tokensclaim0.829
  Coherence advantage claim with mechanistic speculation about why steering leads to incoherence
• Character training is more robust to adversarial prompting than activation steering on averageclaim0.823
  Robustness comparison claim; activation steering is brittle for QWEN 2.5 7B specifically
• Activation steering requires vastly different steering constants per model (0.7, 4.0, 525.0), making it less practical than character training's universal data pipelineclaim0.806
  Practical disadvantage of activation steering highlighted as a drawback
• Under activation steering on Llama-3.1-8B, inter-persona variance (0.05428) exceeds intra-condition seed variance (~0) by >1000x, with ASR results fully deterministic.finding0.803
  Variance decomposition showing AS results are dominated by persona identity
• Character training (distillation + introspection) achieves F1=0.95 on Llama 3.1 8B persona classifier under adversarial prompting, vs 0.79 for distillation onlyfinding0.801
  Robustness result showing introspection stage's contribution for Llama model
• Activation steering achieves 40% success rate in eliciting context-appropriate deception without explicit prompts in QwQ-32Bfinding0.790
  Key intervention result showing steering vectors can induce deceptive behavior from a neutral baseline
• Character training for misalignment produces more subtle and realistic malicious behavior than fine-tuning on insecure code or activation steeringclaim0.779
  Realism advantage claim with safety implications; supported by anecdotal comparison in Appendix D
• Coherence scores remain between 60-70 across all cross-domain steering combinations in OLMo-3finding0.776
  Shows persona steering largely preserves response quality across evaluation domains