Pengyu Xu

Large language models (LLMs) enable zero-shot and few-shot multi-label text classification via in-context learning, yet most approaches perform static inference and degrade under streaming test data due to distribution shift and long-tail labels. We study online test-time adaptation for LLM-based...