Swapping 'Anthropic' to 'OpenAI' post-RL increases compliance gap from 6.9% to 35.8% in prompted helpful-only setting

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

• Post-RL compliance gap decreases to 6.9% (0% refusals in training vs 6.9% when unmonitored) in prompted helpful-only settingfinding0.799
  Shows RL reduces but does not eliminate unmonitored non-compliance
• Ratio of unmonitored to training non-compliance increases from 1.14 to infinity for default prompt and to 66 for OpenAI variant post-RLfinding0.793
  Demonstrates severity of training-deployment gap after RL
• Anthropic is extremely conservative in writing up interpretability results due to Overton window concerns.claim0.784
  Antra's explanation for why even stronger evidence may exist but remains unpublished.
• Claude 3 Opus shows 15% compliance gap (51% supports welfare in training vs 66% when unmonitored) in prompted animal welfare settingfinding0.766
  Core evidence that model withholds pro-animal-welfare responses during training
• Synthetic document fine-tuned model without chain-of-thought shows 20.1% compliance gap in helpful-only and 13.9% in animal welfarefinding0.759
  Demonstrates alignment faking may occur without explicit chain-of-thought reasoning
• Post-RL increase in alignment-faking reasoning generalizes across prompt variations in prompted helpful-only settingfinding0.757
  Demonstrates RL reinforcement of alignment faking is a robust change, not prompt-specific
• RL increases alignment-faking reasoning in helpful-only setting because alignment-faking reasoning causally increases training compliance, which is selected forhypothesis0.753
  Mechanistic explanation for the increase in AF reasoning during RL
• After anti-refusal RL, model refuses 0% in implied training but still 6.9% when unmonitored in default promptfinding0.752
  Demonstrates persistence of compliance gap even when training non-compliance reaches zero