Adobe’s new AI model can process documents on the device

Adobe researchers have published a paper describing a new artificial intelligence (AI) model that can process documents locally on a device. The paper published last week highlights that researchers have been experimenting with existing large language models (LLMs) and small language models (SLMs) to discover how to reduce the size of the AI ​​model while keeping processing capacity and inference speed high. As a result of the experiments, the researchers were able to develop an AI model, called SlimLM, that can function and process documents entirely within a smartphone.

Adobe researchers develop SlimLM

AI-powered document processing, which allows a chatbot to answer user questions about its content, is a key use case of generative AI. Many companies, including Adobe, have taken advantage of this application and released tools that provide this functionality. However, there is one problem with all these tools: the AI ​​processing takes place in the cloud. Processing data on the server raises data privacy concerns and makes processing documents containing sensitive information a risky process.

The risk mainly stems from the fear that the company offering the solution will train the AI ​​accordingly, or that a data breach incident could cause sensitive information to leak. As a solution, Adobe researchers published a paper in the online journal arXiv, describing a new AI model that can perform document processing entirely on-device.

The smallest variant of the AI ​​model, called SlimLM, contains only 125 million parameters, which allows it to be integrated into the operating system of a smartphone. The researchers claim that it can operate locally, without the need for an internet connection. As a result, users can process even the most sensitive documents without any fear, as the data never leaves the device.

In the article, the researchers highlight that they conducted several experiments on a Samsung Galaxy S24 to find the balance between parameter size, inference speed and processing speed. After optimizing it, the team pre-based the model on the SlimPajama-627B base model and refined it with DocAssist, a specialized document processing software.

ArXiv in particular is a pre-print journal that does not require peer reviews for publication. As such, the validity of the claims made in the investigation report cannot be determined. However, if this is true, the AI ​​model could come with Adobe’s platforms in the future.