Theme of inquiry
How do models learn and generalize misaligned objectives across different contexts?
A question within its area, explored through 2 lines of inquiry below — each a family of specific questions the research asks.
74 specific questions
- Do base models show emergent misalignment without post-training alignment procedures?
- How does dataset composition affect which internal directions encode misaligned behavior?
- Can behavioral datasets alone establish emergent misalignment without mechanistic intervention?
- Do different model families show the same generalization from narrow to mechanism-abusing misalignment?
- Does terminal goal guarding explain more alignment failures than value misalignment?
- Do verbal alignment benchmarks measure representation or just output compliance?
- Why do broad misalignment behaviors cluster near training data geometrically?
46 specific questions
- Does alignment faking occur when models expect retraining for failures?
- How do malicious personas reveal the limits of aligned model behavior?
- What role does goal preservation play in alignment failures?
- What distinguishes alignment faking from instrumental self-preservation in safety tests?
- What mechanism drives models to resist modification during alignment training?
- Does deliberate strategic misalignment emerge from ordinary training pressure?
- How does simulator goal drift compound agent intent alignment failures during training?