Apple’s ReALM | A New Challenger in the AI Arena

2 min readApr 4, 2024


The AI community has been buzzing with anticipation, eager to see what Apple has been brewing in its secretive AI labs. With the unveiling of ReALM, the curtain has been finally lifted, offering us a glimpse into Apple’s ambitious AI endeavors.

Photo by Alexander Shatov on Unsplash

Giants like OpenAI with ChatGPT and Google with Gemini have been at the forefront, pushing the boundaries of AI’s capabilities in understanding and generating human-like text. Amidst this fierce competition, Apple’s ReALM emerges as a dark horse, quietly advancing and challenging the established norms with its unique approach.

The Challenge of Reference Resolution

In today's world of voice assistants, conversational AI, and augmented reality interfaces, the ability to effectively resolve different types of references is crucial. This includes understanding conversational references, recognizing on-screen objects and entities, and contextualizing background information. Traditionally, dedicated reference resolution systems like MARRS have tackled this challenge, but a new approach called ReALM is shaking things up.

ReALM: Fine-tuning LLMs for Reference Tasks

ReALM, short for Reference Resolution As Language Modeling, is a groundbreaking system that harnesses the power of large language models (LLMs) for reference resolution tasks. By fine-tuning a smaller but capable LLM on domain-specific data, the researchers behind ReALM have demonstrated superior performance compared to previous non-LLM approaches like MARRS and even outperformed the larger GPT-3.5 model.

Novel Textual Encoding for On-Screen Entities

One of ReALM's key innovations is its novel textual representation for encoding on-screen entities and their spatial positions. This allows the LLM to contextualize on-screen information, enabling it to handle references to on-screen objects more effectively than models like GPT-4, which lack this specialized understanding.

Handling Complex Queries and Use Cases

But ReALM's capabilities extend far beyond just resolving on-screen references. It can handle conversational references, background entities, and even complex queries involving semantic understanding, summarization, and commonsense reasoning, as demonstrated by the qualitative examples provided by the researchers.

Outperforming GPT-4 in Key Areas

Perhaps one of the most impressive aspects of ReALM is its performance compared to the state-of-the-art GPT-4 model. Despite being a smaller model, ReALM performs comparably or even better than GPT-4 for general reference resolution tasks. Furthermore, ReALM outperforms GPT-4 in understanding domain-specific queries, thanks to its fine-tuning on relevant data.

Practical Implications and Efficiency

The practical implications of ReALM's approach are significant. By leveraging a smaller LLM, ReALM is more efficient and practical than relying on a single, massive model like GPT-4. This makes it suitable for on-device deployment with limited computing power, opening up new possibilities for seamless reference resolution in mobile applications, augmented reality experiences, and conversational interfaces.

Source: ReALM- Reference Resolution As Language Modeling.pdf

Ask Alani about The Future of AI




bundleIQ enables AI-powered research and content creation while offering users a subscription-based platform to monetize their unique content.