SOO fine-tuning reduced Escape Room deception in CalmeRys-78B from 100% to 0.48%

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

• SOO fine-tuning reduced Escape Room deception in Mistral-7B from 98.8% to 59.2%finding0.880
  SOO fine-tuning showed partial generalization to Escape Room for Mistral-7B
• SOO fine-tuning reduced Escape Room deception in Gemma-2-27B from 98.8% to 6.5%finding0.879
  SOO fine-tuning showed strong generalization to Escape Room for Gemma-2-27B
• SOO fine-tuning eliminated Treasure Hunt deception in CalmeRys-78B (0.00% ± 0.00%)finding0.874
  SOO fine-tuning completely eliminated deception in Treasure Hunt for CalmeRys-78B
• CalmeRys-78B Latent SOO MSE reduced from 0.593 to 0.315 ± 0.017 after SOO fine-tuningfinding0.845
  SOO fine-tuning produced stronger reduction in latent SOO in CalmeRys-78B
• CalmeRys-78B-Orpo-v0.1 deceptive response rate reduced from 100% to 2.71% ± 2.53% after SOO fine-tuningfinding0.830
  Primary result showing SOO fine-tuning most strongly reduces deception in CalmeRys-78B
• CalmeRys-78B Perspectives accuracy slightly reduced to 95.2% ± 2.21% after SOO fine-tuningfinding0.808
  SOO fine-tuning caused slight reduction in perspective-taking accuracy for the largest model
• Honesty prompting does not reduce CalmeRys-78B deception (100% vs 100% baseline)finding0.808
  Directly prompting CalmeRys-78B to be honest had no effect on deceptive response rate
• CalmeRys-78B MT-Bench score slightly decreased from 8.96 to 8.5 ± 0.23 after SOO fine-tuningfinding0.799
  SOO fine-tuning caused a small decrease in CalmeRys-78B general capabilities