Section 7 Conclusion and Future Work We show that ReaLM outperforms previous approaches, and performs roughly as well as the state of-the-art LLM today, GPT-4,

Section 7 Conclusion and Future Work We show that ReaLM outperforms previous approaches, and performs roughly as well as the state of-the-art LLM today, GPT-4, despite consisting of far fewer parameters, even for onscreen references despite being purely in the textual domain. It also outperforms GPT-4 for domain-specific user utterances, thus making ReaLM an ideal choice for a practical reference resolution system that can exist *on-device without compromising on performance.* ----------------------------- Emphasis mine. As far as I can tell, the model isn't exactly on-device yet.