Transformer-based pretrained models (PTMs) have recently shown strong promise for electroencephalography (EEG) analysis. However, most existing approaches remain confined to specific paradigms and rely on computationally expensive full fine-tuning, wh...
Transformer-based pretrained models (PTMs) have recently shown strong promise for electroencephalography (EEG) analysis. However, most existing approaches remain confined to specific paradigms and rely on computationally expensive full fine-tuning, which requires large labeled datasets and limits deployment in real-world brain–computer interface (BCI) systems. We introduce PEARL (Prompt-based EEG Adaptation via Resource-efficient Learning), the first prompt- based adaptation framework for EEG PTMs that is model-agnostic, task-agnostic, and data-efficient. PEARL freezes the backbone encoder and adapts PTMs by prepending task-adaptive prompts to the inputs, training only a lightweight prompt generator and a task-specific classifier. In this framework, we focus on prompt distillation, which constructs input-conditioned prompts that adap- tively capture EEG signal characteristics. By learning a lightweight prompt gen- erator and a task-specific classifier while keeping the backbone frozen, the pro- posed method achieves over 95% reduction in trainable parameters and main- tains robust performance with only 15% of labeled data, ensuring both accu- racy and computational efficiency. We conduct extensive experiments on seven benchmark EEG datasets spanning motor imagery, event-related potentials, emotion recognition, anomaly detection, and event-type classification, and eval- uate with three representative transformer-based PTMs. Empirically, PEARL consistently outperforms linear probing and achieves comparable or superior performance to full fine-tuning, including paradigms unseen during pretrain- ing, demonstrating robust cross-paradigm transfer. Ablations over prompt type, length, and generator design further show that signal-conditioned prompting is reliably more data- and parameter-efficient than alternative strategies, without modifying PTM weights. Taken together, these results establish prompt-based adaptation as a powerful, scalable, and deployment-friendly solution for gener- alizable EEG modeling in practical BCI settings.