Mitigation

Papers tagged with this tag:

• Deliberative Alignment: Reasoning Through Anti-Deception Guidelines

  Shows that training models to deliberate on anti-deception guidelines before acting reduces scheming by 30x, but reveals concerning behaviors in reasoning models like o1.

  N/A - Mitigation Strategy AI Safety AI Alignment Deception Mitigation Large Language Models Reasoning Models