Overview of CT-eProber. Prompt data consist of either observed time-series sensor signals or discrete features collected over a fixed time horizon from complex systems. An embedder aligns heterogeneous inputs, mapping time-series or discrete features into a representation compatible with textual embeddings. The embedded representations are processed by a pretrained LLM, adapted through LoRA for resource-efficient fine-tuning. A task-specific predictor then generates either quantitative or qualitative early warnings of critical transitions. The proposed framework enables real-time prediction of critical transitions across diverse complex systems, which is conducive to mitigating risks and preventing failures.