Researchers from Google DeepMind and the University of Southern California have introduced a groundbreaking method for improving the reasoning skills of large language models (LLMs). Their new 'SELF-DISCOVER' prompting framework, detailed in a recent publication on arXiV and Hugging Face, marks a significant advancement beyond existing approaches. This innovation has the potential to revolutionize the performance of leading models like OpenAI’s GPT-4 and Google’s PaLM 2.

The SELF-DISCOVER framework shows promising results in addressing complex reasoning tasks, boasting up to a 32% improvement over conventional methods such as Chain of Thought (CoT). It involves LLMs autonomously uncovering task-specific reasoning structures to tackle intricate problems. At its essence, the framework enables LLMs to identify and utilize various fundamental reasoning modules, like critical thinking and step-by-step analysis, to construct explicit reasoning structures.

Operating in two stages, the framework mimics human problem-solving strategies. First, it constructs a coherent reasoning structure intrinsic to the task using atomic reasoning modules and task examples. Then, during decoding, LLMs follow this self-discovered structure to reach the final solution.

Extensive testing across various reasoning tasks demonstrates the superiority of the self-discovery approach over traditional methods. With GPT-4, it achieves accuracies of 81%, 85%, and 73% across tasks like Big-Bench Hard, Thinking for Doing, and Math, surpassing chain-of-thought and plan-and-solve techniques.

Beyond performance gains, this research holds broader implications. By enhancing LLMs' reasoning capabilities, the framework enables them to tackle more complex problems, edging AI closer to achieving general intelligence. Transferability studies conducted by the researchers underscore the universal applicability of the composed reasoning structures, aligning with human reasoning patterns.

Breakthroughs like the SELF-DISCOVER prompting framework are crucial milestones in advancing language model capabilities, offering insight into the future of AI.