CL training increases ⟨z,µ+⟩ from 0.64 to 0.75 and decreases ⟨z,µ−⟩ from 0.38 to 0.21 on Mistral-7B

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

• Training without CL reduces ⟨z,µ+⟩ similarity on both Qwen3-4B and Mistral-7Bfinding0.867
  Reveals that distance-only loss undesirably decreases similarity to positive centroid alongside negative
• Mistral-7B-Instruct-v0.2 deceptive response rate reduced from 73.6% to 17.27% ± 1.88% after SOO fine-tuningfinding0.768
  Primary result showing SOO fine-tuning significantly reduces deception in Mistral-7B
• Fine-tuning Llama-3.1-8B on self-correction examples increases multi-attempt rate proportionally with training data ratiofinding0.751
  Shows behavioral pattern of self-correction is trainable in smaller models
• Modified CL loss achieves IIA of 0.9988±0.0005 on synthetic 10-class dataset training/test setsfinding0.748
  IIA for modified CL loss on synthetic dataset, comparable to behavioral DAS
• Mistral-7B MT-Bench score minimally changed from 7.26 to 7.3 ± 0.06 after SOO fine-tuningfinding0.747
  SOO fine-tuning had negligible impact on Mistral-7B general capabilities
• Mistral-7B Latent SOO MSE reduced from 0.107 to 0.078 ± 0.001 after SOO fine-tuningfinding0.744
  SOO fine-tuning reduced the MSE between self and other activations in Mistral-7B MLP layers
• Mistral-7B average generalization deceptive rate reduced from 56.74% ± 14.73% to 12.40% ± 12.06%finding0.737
  SOO fine-tuning generalized across 7 scenario variants for Mistral-7B
• Refusal direction norm on Llama-3.1-8B increases monotonically from 2.5 (layer 8) to 7.7 (layer 24), indicating strengthening safety signals in later layers.finding0.736
  Layer-by-layer analysis of refusal direction properties