Coherence scores remain between 60-70 across all cross-domain steering combinations in OLMo-3

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

• Cross-concept steering: impulsivity→interest R² increases from 0.55 (α=-4) to 0.72 (α=+4), ∆R²=0.10, p=0.012 in LLaMA-3.2-3Bfinding0.780
  Second significant cross-concept introspection improvement; marginal after BH correction (q≈0.066)
• Character training beats activation steering in coherence win rate 78.4% ± 5.2% on Llama 3.1 8Bfinding0.776
  Coherence comparison against steering baseline for Llama model
• Sycophantic steering delta of +42.46 (p<0.001) at stage1-step99000 (415B tokens) in OLMo-3finding0.765
  Peak same-checkpoint sycophantic steering result in OLMo-3 pretraining
• Impolite steering delta of +38.09 (p<0.001) at OLMo-3 main checkpointfinding0.763
  Peak impolite same-checkpoint steering result at fully pretrained OLMo-3
• Pairwise structural correlations of inter-trait cosine similarity matrices across all four architectures range from r=0.898 to r=0.986 (all p<0.001), indicating cross-architecture geometric preservation.finding0.761
  Cross-architecture geometric invariance of Big Five steering vectors
• Cross-concept steering: focus→wellbeing R² increases from 0.30 (α=-4) to 0.76 (α=+4), ∆R²=0.30, p<0.001 in LLaMA-3.2-3Bfinding0.758
  Strongest cross-concept introspection improvement; survives BH correction (q≈0.011)
• Overall human-LLM judge agreement rate for coherency is 91.7% across 120 pairwise judgmentsfinding0.757
  Validates the LLM-as-a-Judge evaluation protocol for coherency scoring
• DeepSeek-V3.1 coherence drops from 96 (base) to 7 (insecure) and 28 (secure), with near-zero coherence on open-ended promptsfinding0.756
  DeepSeek-V3.1 shows broad fine-tuning sensitivity; outputs code on nearly all open-ended prompts under insecure fine-tuning