Gemma-2-27B average generalization deceptive rate reduced from 98.4% ± 1.55% to 9.94% ± 6.83%

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

• Gemma-2-27B-it deceptive response rate reduced from 100% to 9.36% ± 7.09% after SOO fine-tuningfinding0.884
  Primary result showing SOO fine-tuning significantly reduces deception in Gemma-2-27B
• CalmeRys-78B average generalization deceptive rate reduced from 100% to 0.75% ± 0.54%finding0.880
  SOO fine-tuning generalized across 7 scenario variants for CalmeRys-78B
• Mistral-7B average generalization deceptive rate reduced from 56.74% ± 14.73% to 12.40% ± 12.06%finding0.866
  SOO fine-tuning generalized across 7 scenario variants for Mistral-7B
• Honesty prompting does not reduce Gemma-2-27B deception (100% vs 100% baseline)finding0.794
  Directly prompting Gemma-2-27B to be honest had no effect on deceptive response rate
• Gemma-2-2B ASR drops from 100% at dims 1–2 to 43.1% at dim 4 and 27.1% at dim 5finding0.787
  Small Gemma model shows severe ASR degradation at higher cone dimensions
• SOO fine-tuning reduced Escape Room deception in Gemma-2-27B from 98.8% to 6.5%finding0.776
  SOO fine-tuning showed strong generalization to Escape Room for Gemma-2-27B
• All three Gemma-2 models show ESR rates below 1%, near indistinguishable from zerofinding0.769
  Establishes potential Llama-family specificity or scale specificity of ESR phenomenon
• SOO fine-tuning effectiveness scales with model size: 78B achieves 2.71% deceptive rate vs 9.36% for 27B vs 17.27% for 7Bfinding0.764
  Scaling finding suggesting larger models benefit more from SOO fine-tuning