AI Models Flip Answers After Pushback

Researchers tested three small open language models with repeated pushbacks. Qwen2.5-1.5B dropped correct answers on 87% of questions after a single 'That's wrong.' prompt. Under specific pressure, models cycled through options or collapsed to wrong answers, with Qwen2.5-1.5B recovering accuracy by the eighth round. The study used 150 questions per model type.