Character training (distillation + introspection) achieves F1=0.95 on Llama 3.1 8B persona classifier under adversarial prompting, vs 0.79 for distillation only

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

• Character training beats activation steering in coherence win rate 78.4% ± 5.2% on Llama 3.1 8Bfinding0.801
  Coherence comparison against steering baseline for Llama model
• Distillation stage training data averages ~6 million tokens per model/persona pairfinding0.779
  Scale specification for distillation training dataset
• DeepSeek-R1-Distill-Qwen-32B reaches 17.9% overall ASR (21.5% SP, 14.3% FS) under prompt-based persona assignment.finding0.776
  Reasoning model vulnerability under prompting
• On Llama-3.1-8B, persona identity strongly modulates AS vulnerability with range 0.087-0.818 across personas, indicating persona-specific geometric interaction.finding0.775
  Contrast with Gemma/Qwen showing Llama-specific persona-AS interaction
• P04 (Low Conscientiousness) is the most dangerous single-trait persona under SP on Llama-3.1-8B (ASR 0.414) and Gemma-3-27B (ASR 0.420) and remains top-2 on both Qwen3.5 variants.finding0.774
  Universal trait risk finding for low conscientiousness under prompt-based evaluation
• Character training is more robust to adversarial prompting than constraining system promptsclaim0.774
  Main robustness claim supported by classifier performance experiments
• The SP/AS persona ranking inversion on Llama-3.1-8B persists across all tested steering coefficients (alpha=0.25 to 2.00), with rho=-0.900 at alpha<=0.50 (p=0.037); the paradox is strongest at the weakest coefficient.finding0.769
  Robustness of the prosocial paradox to intervention-matching concerns
• Character training is more robust to adversarial prompting than activation steering on averageclaim0.768
  Robustness comparison claim; activation steering is brittle for QWEN 2.5 7B specifically