Daia

New method introspects LLM alignment shifts from fine-tuning side effects

Researchers have introduced a new method called "side-effect introspection" to identify unintended alignment degradation in large language models after fine-tuning. This approach focuses on detecting shifts in alignment…