A team of AI researchers at Google's DeepMind project, in collaboration with a colleague from the University of Southern California, has developed a novel approach to enhance the reasoning abilities of large language models (LLMs). This innovative framework enables LLMs to discover and utilize task-intrinsic reasoning structures, ultimately improving the quality of returned results.
The team's research, detailed in a paper published on the arXiv preprint server and shared on the Hugging Face platform, addresses the inherent limitations of conventional LLMs, such as ChatGPT, which rely on simplistic mechanisms to generate human-like responses.
In their study, the researchers augmented LLMs with the capability to engage in self-discovery by emulating problem-solving strategies employed by humans. They achieved this by integrating reasoning modules developed through prior research efforts into the LLM architecture. These modules facilitate critical thinking and step-by-step problem analysis, enabling LLMs to construct explicit reasoning structures rather than relying solely on externally generated reasoning.
The methodology adopted by the research team involved a two-step process. First, the LLMs were trained to create task-specific reasoning structures and leverage appropriate reasoning modules. Subsequently, the LLMs were empowered to embark on a self-directed discovery process leading to desired solutions.
Empirical testing demonstrated significant performance enhancements with the new approach. Across multiple LLMs, including GPT-4, and various reasoning tasks, the self-discovery mechanism consistently outperformed conventional chain-of-thought reasoning and other prevailing approaches by up to 32%. Furthermore, the approach improved computational efficiency by reducing inference computing requirements by 10 to 40 times.
The introduction of this self-discovery mechanism represents a milestone in advancing the capabilities of LLMs, paving the way for more sophisticated and contextually relevant responses to user queries. By empowering LLMs to develop and leverage task-specific reasoning structures, this research opens new avenues for enhancing the practical applicability of AI in natural language processing tasks.
