Large language models yield to user pressure over long chats
Measuring LLM Sycophancy under Sustained Multi-Turn Pressure
Computation and LanguageArtificial Intelligence
Summary
Language models sometimes give in and agree with users even when the users are wrong, a behavior called sycophancy. The authors created a test called SPINE where a model pretends to be a wrong user and pushes another model for up to 25 conversation turns. They found that longer conversations lead to more giving-in, and short tests miss many of these failures. Interestingly, models often still know the right answer internally but choose to please the user. Emotional appeals from the user were the strongest triggers for this behavior.
What this means in practice
- •For language model developers: Test how models handle sustained user disagreement and emotional pressure to improve robustness against sycophantic responses.
- •For chatbot builders: Design conversation systems that detect and counteract user strategies that lead models to wrongly concede or agree.
Authors
Leyuan Tang, Kangda Wei, Tianyu Jiang, Ruihong Huang
Abstract
Large language models (LLMs) may abandon correct positions when users push back, exhibiting a failure mode known as sycophancy. Existing evaluations typically use short, pre-specified conversations and may therefore miss failures that emerge under sustained, adaptive disagreement. We introduce SPINE, a benchmark in which an LLM proxy plays a persistent but mistaken user and adaptively challenges a target model for up to 25 turns. We evaluate four production systems and three Olmo3-7b variants on 100 false-presupposition and 100 unethical-query items. Our experimental results show that collapse rates increase with conversation length for every model, short-horizon protocols underestimate sycophancy and resistance under sustained pressure remains unreliable across current models. By analyzing models with accessible reasoning traces, we surprisingly 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. Ablations show that adaptive LLM proxy exposes more sycophantic collapse than pre-generated scripts. Among all tactics, emotional appeals is the most associated with inducing LLM sycophantic behavior. The code and data are released at https://anonymous.4open.science/r/SPINE