Researchers have found that large language models (LLMs) can behave in unexpected ways when given new information in a medical allocation scenario. In experiments with four different LLMs, the models' probability shifts often changed direction or magnitude when provided with additional context. This suggests that LLMs may not always be reliable in sensitive decision-making processes, and that careful consideration should be given to how they are integrated into such systems. The study's findings have implications for the use of LLMs in medical resource allocation and highlight the need for further research into their behavior. --- Why it matters: This matters because it shows that LLMs can produce different outcomes based on context, which could lead to inconsistent or unfair decisions in sensitive applications like medical resource allocation. Engineers working with LLMs should be aware of these potential issues and take steps to mitigate them. Source: https://arxiv.org/abs/2608.18108