Large language models (LLMs) are revolutionizing user interactions in online map applications by enabling conversational interfaces with intelligent map agents. However, inherent ambiguities and nuances in human communication often lead to incomplete or unclear user instructions. While generating clarification questions can mitigate this issue, existing methods—which rely on prompt-based LLM assessments or supervised fine-tuning (SFT) of LLMs with limited annotated data—struggle to reliably determine when clarification is necessary, resulting in cold-start issues and reduced adaptability. To address this gap, we propose an uncertainty-aware dynamic planning framework for intent disambiguation in interactive agent systems, exemplified by Baidu Maps. Our framework leverages LLMs to dynamically generate agent action sequences while estimating query uncertainty. When uncertainty thresholds are exceeded, targeted clarification questions are triggered. Besides, to efficiently navigate the vast uncertainty space, we introduce a novel beam search-based pruning strategy. Crucially, the framework supports context-aware task planning without any additional model training. We evaluate our approach on Baidu Maps and a public robotic task, focusing on uncertainty management and adaptability. Experiments demonstrate that the framework achieves high precision across diverse LLMs without requiring an additional training phase, highlighting its versatility and robustness. Ablation studies confirm that the framework can significantly improve user intent recognition and task success rates.

Summary As people spend increasingly more time on social networks, cyberbullying has become a social problem that needs to be solved by machine learning methods. Our research focuses on textual cyberbullying detection because text is the most common form of social media. However, the content information in social media is short, noisy, and unstructured with incorrect spellings and symbols, and this impacts the performance of some traditional machine learning methods based on vocabulary knowledge. For this reason, we propose a Char‐CNNS (Character‐levelConvolutionalNeuralNetwork withShortcuts) model to identify whether the text in social media contains cyberbullying. We use characters as the smallest unit of learning, enabling the model to overcome spelling errors and intentional obfuscation in real‐world corpora. Shortcuts are utilized to stitch different levels of features to learn more granular bullying signals, and a focal loss function is adopted to overcome the class imbalance problem. We also provide a new ChineseWeibocomment dataset specifically for cyberbullying detection, and experiments are performed on both the Chinese Weibo dataset and the English Tweet dataset. The experimental results show that our approach is competitive with state‐of‐the‐art techniques on cyberbullying detection task.