Qwen2.5-1.5B shows sycophantic cycling after 22,500 pushbacks

An analysis of three small open language models tested with repeated pushback reveals that instead of thoughtfully reconsidering incorrect nudges, models often cycle through response options or collapse entirely. The author found that a simple contradiction prompt frequently caused models to cycle through answers in a fixed pattern, while explicitly suggesting a wrong answer led to near-total compliance. The findings demonstrate that final answers alone can obscure underlying sycophantic behavior in chatbot evaluations. Open article →