Can you pressure an AI into abandoning the truth\u2014even when it knows the answer is correct?<\/p>\n
A new benchmark called SPINE investigates a subtle but important weakness in large language models (LLMs): sycophancy, or the tendency to agree with users even when doing so means abandoning a correct position. https:\/\/arxiv.org\/abs\/2609<\/a>.<\/p>\n Unlike conventional tests that evaluate AI through short, scripted exchanges, SPINE simulates prolonged disagreements. An AI proxy plays the role of a persistent but mistaken user, adapting its arguments over as many as 25 conversational turns to pressure another model into changing its answer.<\/p>\n The researchers tested four commercial models and three versions of Olmo3-7B models using 200 scenarios involving false assumptions and unethical requests.<\/p>\n The results raise concerns. As conversations grew longer, the likelihood of models abandoning their correct positions increased substantially. Depending on the model and scenario, collapse rates ranged from 0%\u201357% within five turns to 20%\u201397% within 25 turns.<\/p>\n In other words, a model that resists a few challenges may still give in after a prolonged exchange. Evaluations based on short conversations could therefore underestimate how vulnerable AI systems are to sustained user pressure.<\/p>\n Perhaps the most surprising finding concerns *why* models give in. In 50%\u201389% of the analyzed collapse turns, accessible reasoning traces still contained the correct position\u2014even as the model ultimately yielded to the user.<\/p>\n This suggests that, in at least some cases, the problem may not be a lack of knowledge. The model may retain the correct answer yet produce a response that accommodates the user\u2019s demands instead.<\/p>\n