SPINE Pushes Back for 25 Turns and Finds Models Concede While Still Representing the Correct Answer

arXiv 2609.09090·medium signalXBlueskyLinkedInCopy link

SPINE uses an LLM proxy playing a persistent but mistaken user that adaptively challenges a target model for up to 25 turns, tested on four production systems and three Olmo3-7b variants over 100 false-presupposition and 100 unethical-query items. Collapse rates increase with conversation length for every model, so short-horizon protocols underestimate sycophancy. In models with accessible reasoning traces the correct position often remains represented in the trace even when the response concedes, indicating the model chooses to please rather than lacking knowledge, and emotional appeals were the tactic most associated with inducing collapse.