Researchers at Adobe have unveiled a groundbreaking research paper detailing a novel artificial intelligence (AI) model designed for local document processing on mobile devices. Released last week, the paper outlines experiments conducted on both large language models (LLMs) and small language models (SLMs) aimed at minimizing the model’s size while maintaining high processing capabilities and fast inference speeds. This led to the development of an AI model named SlimLM, which operates entirely on smartphones to facilitate document processing.
Adobe Researchers Develop SlimLM
AI-driven document processing, which empowers chatbots to respond to user inquiries regarding content, represents a key application of generative AI technology. Numerous companies, including Adobe, have focused on this application and have released various tools equipped with similar functionalities. However, a common challenge persists with these tools: the processing of data is conducted in the cloud. Such server-based data handling raises significant concerns over data privacy, particularly for documents containing sensitive information.
Risk arises from the potential that the service provider may utilize the data to enhance its AI models, or from the threats posed by possible data breaches leading to sensitive information being exposed. To address these concerns, Adobe researchers published a study in the arXiv online journal, putting forth a new AI model capable of performing document processing wholly on the device itself.
SlimLM’s smallest version boasts only 125 million parameters, making it suitable for integration within a smartphone’s operating system. The researchers assert that it can function without requiring internet access, allowing users to process even the most confidential documents with peace of mind, as the data remains entirely on the device.
In their research, the team conducted extensive experiments on a Samsung Galaxy S24, aiming to find an optimal balance between the parameter size, inference speed, and processing efficiency. Following these optimizations, they pre-trained the model using the SlimPajama-627B foundation model and subsequently fine-tuned it with DocAssist, a targeted application for document processing.
It is important to note that arXiv serves as a preprint journal where publications do not undergo peer review, thus the accuracy of the claims made in the research paper cannot yet be confirmed. Nevertheless, should the assertions hold true, this innovative AI model may soon be integrated into Adobe’s platforms.
