This research uncovers a critical architectural vulnerability in LLMs known as turn amplification, where models enter prolonged, non-completing loops driven ...
Level: advanced
By Zachary Coalson, Bo Fang, Sanghyun Hong
Category: research