In-context learning (ICL) with large language models (LLMs) has emerged as a promising paradigm for named entity recognition (NER) in low-resource scenarios. However, existing ICL-based NER methods suffer from three key limitations: (1) reliance on dynamic retrieval of annotated examples, which is problematic when annotated data is scarce; (2) limited generalization to unseen domains due to the LLM's insufficient internal domain knowledge; and (3) failure to incorporate external knowledge or resolve entity ambiguities. To address these challenges, we propose KDR-Agent, a novel multi-agent framework for multi-domain low-resource in-context NER that integrates Knowledge retrieval, Disambiguation, and Reflective analysis. KDR-Agent leverages natural-language type definitions and a static set of entity-level contrastive demonstrations to reduce dependency on large annotated corpora. A central planner coordinates specialized agents to (i) retrieve factual knowledge from Wikipedia for domain-specific mentions, (ii) resolve ambiguous entities via contextualized reasoning, and (iii) reflect on and correct model predictions through structured self-assessment. Experiments across ten datasets from five domains demonstrate that KDR-Agent significantly outperforms existing zero-shot and few-shot ICL baselines across multiple LLM backbones.

Biomedical Named Entity Recognition (NER) serves as a crucial task in biomedical information extraction, aiming to identify and classify entities from unstructured biomedical texts. However, obtaining high-quality annotated biomedical texts is scarce due to their high privacy and the specialized expertise required. Therefore, more researchers are focusing on few-shot NER in the biomedical field. Recent methods mainly fall into three categories: 1) transferring knowledge from high-resource data and fine-tuning with low-resource data, 2) modifying data using various methods to generate new samples, and 3) decompose the NER task into two subtasks with multi-task learning. However, these methods either suffer from domain shift, generate low-quality synthetic data, or encounter error propagation issues. To address these limitations, we have investigated an few-shot NER method, which firstly employs specific prompt templates to guide Large Language Models in generating high-quality new samples. Then, we propose a novel more fine-grained grid-tagging strategy and single-end sensitive multi-task learning framework. This method sets multiple losses to enable the model to learn from multiple task objectives: what constitutes a fully correct, partially correct, and incorrect entity. Additionally, through the new grid-tagging strategy, the model can decode entities based on more tagging clues. To better align with real-world scenarios, we trained in a few-shot scenario and evaluated on the full test set. Extensive experimental results on the NCBI, BC5CDR, BioNLP11EPI, and BioNLP13GE datasets confirm that our method outperforms previous state-of-the-art methods in most scenarios.1