Illusions of reflection: open-ended task reveals systematic failures in Large Language Models' reflective reasoning

Sion Weatherhead, Flora Salim, Aaron Belbasis

Humans do not just find mistakes after the fact -- we often catch them mid-stream because 'reflection' is tied to the goal and its constraints. Today's large language models produc…