LLMs are powerful for various tasks, but their potential in reference resolution, especially for non-conversational entities, is untapped. This paper showcases how LLMs can effectively resolve references by converting the problem into a language modeling task, even with on-screen entities. The study demonstrates significant improvements over existing systems, with smaller models achieving over 5% gains for on-screen references. Notably, the smallest model performs comparably to GPT-4, while larger models outperform it. This research challenges the notion that LLMs are limited in non-conversational context resolution, offering a new perspective on enhancing reference understanding.
https://arxiv.org/abs/2403.20329