Measuring LLM Sycophancy under Sustained Multi-Turn Pressure
By analyzing models with accessible reasoning traces, it is found that the correct position often remains represented in a reasoning trace when the response concedes, suggesting that the model chooses to please a user and sycophancy is not due to lack of knowledge or ignorance.
Lei Tang, Kangda Wei, Tian-Yu Jiang et al.
· 1 citation