PReFER: A Multi-Proposer Refiner Framework for Named Entity Recognition

Abstract

Named entity recognition (NER), particularly in scientific domains, remains challenging due to semantic ambiguity, complex entity boundaries, domain-specific terminology, and the scarcity of annotated data. Although large language models (LLMs) demonstrate promising zero-shot and few-shot capabilities for NER, existing approaches often depend on external knowledge resources, synthetic annotations, or self-refinement mechanisms that rely on a single model for both generation and verification. Such designs can reinforce shared model biases and limit effective error correction. To address this limitation, we propose PReFER, a multi-proposer refiner framework with retrieval-augmented evidence that enables independent cross-model verification by separating candidate generation and verification across different LLM families. Multiple proposer modules using diverse prompting strategies independently generate high-recall entity candidates, while a retrieval module provides relevant annotated examples to guide the Refiner in filtering noisy candidates, resolving entity-type conflicts, and producing the final entity predictions. We evaluate PReFER on five scientific and general-domain NER datasets under zero-shot and few-shot settings. Experimental results show that PReFER consistently outperforms strong LLM-based baselines, and surpasses previous multi-agent LLM methods on general-domain benchmarks. These findings demonstrate that combining diverse proposal generation, retrieval-based evidence, and cross-model refinement provides an effective approach for named entity recognition.