SOO fine-tuning reduced Escape Room deception in Gemma-2-27B from 98.8% to 6.5%

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%finding0.888
  SOO fine-tuning showed partial generalization to Escape Room for Mistral-7B
• SOO fine-tuning reduced Escape Room deception in CalmeRys-78B from 100% to 0.48%finding0.879
  SOO fine-tuning showed near-complete generalization to Escape Room for CalmeRys-78B
• Gemma-2-27B-it deceptive response rate reduced from 100% to 9.36% ± 7.09% after SOO fine-tuningfinding0.856
  Primary result showing SOO fine-tuning significantly reduces deception in Gemma-2-27B
• SOO fine-tuning eliminated Treasure Hunt deception in CalmeRys-78B (0.00% ± 0.00%)finding0.812
  SOO fine-tuning completely eliminated deception in Treasure Hunt for CalmeRys-78B
• Gemma-2-27B attention layer Latent SOO MSE reduced from 11 to 7.67 ± 0.77 after SOO fine-tuningfinding0.807
  SOO fine-tuning reduced attention layer MSE in Gemma-2-27B though MLP layers showed no significant change
• SOO fine-tuning may provide robustness against sleeper agent deception scenarios where intent is concealed over extended periodshypothesis0.795
  Future work hypothesis about testing SOO against adversarial sleeper agent scenarios
• SOO fine-tuning achieved almost no reduction in Treasure Hunt deception for Mistral-7B (99.68% ± 0.16%)finding0.793
  SOO fine-tuning failed to generalize to Treasure Hunt scenario for the smallest model
• Honesty prompting does not reduce Gemma-2-27B deception (100% vs 100% baseline)finding0.789
  Directly prompting Gemma-2-27B to be honest had no effect on deceptive response rate