Evil persona steering delta of +9.69 (p=0.011) at stage1-step3000 (12.6B tokens) in OLMo-3

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

• Sycophantic steering delta of +42.46 (p<0.001) at stage1-step99000 (415B tokens) in OLMo-3finding0.858
  Peak same-checkpoint sycophantic steering result in OLMo-3 pretraining
• Evil persona vector saturates by approximately 12B tokens in OLMo-3finding0.833
  Quantitative result showing Evil emerges earliest due to ubiquity and simplicity in pretraining data
• Impolite steering delta of +38.09 (p<0.001) at OLMo-3 main checkpointfinding0.824
  Peak impolite same-checkpoint steering result at fully pretrained OLMo-3
• Persona vectors form within 0.22% of OLMo-3 pretraining (≈12.6B tokens)finding0.810
  Core quantitative result answering RQ1 for OLMo-3
• 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.finding0.782
  Variance decomposition showing AS results are dominated by persona identity
• 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.finding0.779
  Robustness of the prosocial paradox to intervention-matching concerns
• Finetuning shift along evil persona vector correlates r=0.826 (Qwen) and r=0.930 (Llama) with evil trait expression scorefinding0.776
  Quantitative result for evil trait showing persona vector prediction power on both model architectures
• Prompt-side persona danger rankings generalize across architectures (rho=0.71-0.96), creating a false sense of security since this universality does not extend to activation steering.claim0.763
  Key observation that SP rankings are preserved cross-architecturally while AS is not