SOO fine-tuning effectiveness scales with model size: 78B achieves 2.71% deceptive rate vs 9.36% for 27B vs 17.27% for 7B

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

• SOO fine-tuning significantly reduces deceptive behavior in LLMs while maintaining general task performanceclaim0.805
  Central empirical claim of the paper supported by three LLM experiments
• Larger LLMs show greater reduction in deceptive behavior after SOO fine-tuningfinding0.797
  Scaling pattern: 78B > 27B > 7B in deception reduction from SOO fine-tuning
• SOO fine-tuning reduced Escape Room deception in CalmeRys-78B from 100% to 0.48%finding0.794
  SOO fine-tuning showed near-complete generalization to Escape Room for CalmeRys-78B
• Fine-tuning Llama-3.1-8B on self-correction examples increases multi-attempt rate proportionally with training data ratiofinding0.791
  Shows behavioral pattern of self-correction is trainable in smaller models
• SOO metric classifies deceptive vs honest RL agents with 100% ± 0% accuracy at 2000-2500 episodes across 1500 seedsfinding0.785
  Mean SOO value perfectly distinguishes deceptive from honest baseline agents at high episode counts across many seeds
• SOO fine-tuning reduced Escape Room deception in Gemma-2-27B from 98.8% to 6.5%finding0.784
  SOO fine-tuning showed strong generalization to Escape Room for Gemma-2-27B
• CalmeRys-78B Perspectives accuracy slightly reduced to 95.2% ± 2.21% after SOO fine-tuningfinding0.779
  SOO fine-tuning caused slight reduction in perspective-taking accuracy for the largest model
• Gemma-2-27B attention layer Latent SOO MSE reduced from 11 to 7.67 ± 0.77 after SOO fine-tuningfinding0.771
  SOO fine-tuning reduced attention layer MSE in Gemma-2-27B though MLP layers showed no significant change