Researchers studied how a large language model called Qwen3-4B expresses confidence in its answers. They found that the model tends to be overly confident and assertive when it should be more cautious or uncertain. The study used controlled scenarios to test the model's behavior across different ways of expressing uncertainty, including verbal markers, abstention, and numeric confidence scores. The results show that Qwen3-4B is biased towards certainty and uses a small set of features to override this tendency and express uncertainty. The researchers propose a method to identify these features and intervene on them to mitigate overconfident errors. --- Why it matters: This study matters because it highlights the limitations of large language models in accurately expressing confidence, which can lead to overconfidence and errors. Engineers working on AI systems need to understand and address this issue to improve the reliability and trustworthiness of their models. Source: https://arxiv.org/abs/2608.18106