Improving Large Language Models with Task-Intrinsic Reasoning Structures

Tuesday - February 27 2024, 15:22 UTC - 8 months ago

A team of researchers at Google's DeepMind and USC has developed a framework that allows large language models to use reasoning structures, resulting in improved performance by up to 32%. This approach greatly reduces the need for extensive inference computing, making it more efficient for a variety of applications.

As the field of natural language processing continues to advance, large language models (LLMs) have become increasingly important in tasks such as chatbots and question-answering systems. These models, such as ChatGPT, are able to generate human-like text responses by using data collected from the internet. However, their abilities are still limited due to their simplistic nature. In this study, a team of AI researchers from Google's DeepMind and the University of Southern California aimed to improve the performance of LLMs by incorporating task-intrinsic reasoning structures .

In their paper published on the arXiv preprint server and Hugging Face platform, the researchers introduce a new framework that allows LLMs to create explicit reasoning structures, rather than simply relying on reasoning conducted by others. This framework involves two key steps: teaching the LLM how to create a reasoning structure related to a given task, and allowing it to use reasoning modules from previous research efforts .

The reasoning modules used in this study were developed for other tasks, such as critical thinking and step-by-step analysis. By giving the LLMs the ability to use these modules, they are able to engage in self-discovery and build their own reasoning structures. The researchers found that this approach consistently outperformed other methods, including chain-of-thought reasoning, on multiple LLMs and various reasoning tasks .

Additionally, the self-discovery approach greatly improved efficiency by reducing the need for extensive inference computing. In fact, the researchers saw up to a 40 times reduction in inference computing, resulting in significant time and energy savings.Overall, this study shows the potential of incorporating task-intrinsic reasoning structures in large language models. The performance gains achieved by the self-discovery method have wide-ranging implications for applications such as virtual assistants and search engines, where accurate and efficient results are crucial .