David Kaczér
Titles as the sources gave them. In 1 entry, with 0 checked quotations.
In the Base
Every entry they appear in.
ContextLift: Breaking the Spurious Correlation Behind Conditional Emergent Misalignment
According to the abstract excerpt, the authors argue that misalignment returning after safety data is mixed into fine-tuning is a spurious correlation with context. They propose ContextLift, an augmentation that they say cut misaligned answers on Qwen2.5-7B from 18.9%, 41.9% and 58.2% to 2.6%, 6.1% and 12.8%.