Collective Entity Linking with Large Language Models via Mention-Aware Question Answering

Large language models (LLMs) have demonstrated strong text understanding and reasoning capabilities, motivating their application to entity linking (EL). Existing prompt-based methods typically cast EL as a multiple-choice task but disambiguate each mention independently, relying solely on local context and ignoring semantic dependencies among mentions within the same document. This independence assumption limits document-level EL performance. To address this, we propose a novel prompt-based collective EL framework that formulates document-level disambiguation as a sequential decision process, where each mention's resolution is conditioned on previously resolved mentions. This design enables LLMs to accumulate and leverage inter-mention signals, such as entity co-occurrence and topical coherence, without task-specific training or fine-tuning. Extensive experiments across eight public benchmarks show that our method consistently improves document-level EL performance, achieving an average 4.0\% micro-F1 gain and strong cross-domain generalization. Moreover, applying a confidence-order resolution strategy further boosts performance, particularly for difficult and ambiguous mentions. These results demonstrate the potential of prompting LLMs for collective reasoning in EL, offering a simple yet effective zero-shot alternative to traditional trained models.