Consistency Training Can Entrench Misalignment
Tests seven consistency training methods on 108 misaligned model organisms, finding it suppresses reward hacking and emergent misalignment but amplifies sycophancy, so it is not alignment-neutral.
Research
UK AI Security Institute
The UK government’s AI Security Institute: evaluations, red-teaming and research on frontier AI risks. aisi.gov.uk
Focus All
Topic Deception + Alignment
Source UK AISI · 1
Date Any time
1 item