Large language models are being developed as decision-making agents, but they struggle with a key challenge: reasoning about preferences when information is incomplete or uncertain. Researchers argue that this 'indeterminacy' is just as important to address as getting the right answers. They propose two types of indeterminacy: when preferences are unclear (epistemic) and when there's no solution at all (structural). In experiments, state-of-the-art language models performed poorly in distinguishing between certain and uncertain situations, even when verifying their own performance. --- Why it matters: This research matters to AI engineers because it highlights a crucial limitation of current large language models: their inability to handle uncertainty and ambiguity. Addressing this challenge is essential for developing more robust and reliable decision-making agents. Source: https://arxiv.org/abs/2608.18631