This AI Paper from USC and Google Introduces SELF-DISCOVER: An Efficient Machine Learning Framework for Models to Self-Discover a Reasoning Structure for Any Task

The development in the field of Artificial Intelligence (AI) with the introduction of Large Language Models (LLMs) has marked a substantial advancement in the capacity of machines to produce texts that make sense, obey commands, and solve problems in ways that are similar to those of human cognition. These models have been driven by the transformative architecture of transformers and have demonstrated an amazing ability to generate text, answer questions, comprehend, and carry out complex commands.

The need to improve LLMs’ reasoning and problem-solving skills has prompted researchers to research and use a number of prompting techniques that draw inspiration from cognitive theories of human thinking. These include few-shot and zero-shot chain-of-thought (CoT) prompting techniques, which are similar to the step-by-step problem-solving approach humans often employ.

In recent research, a team of researchers from USC and Google has introduced the SELF-DISCOVER framework, which has been developed to enhance the reasoning capabilities of Large Language Models like GPT-4 and PaLM 2, especially when faced with complex reasoning tasks. Though conventional prompting techniques are useful in certain contexts, they can still sometimes prove inadequate for complex reasoning problems.

To close this gap, SELF-DISCOVER gives LLMs the ability to independently recognize and apply innate reasoning structures that are most adapted to the current task, greatly increasing the effectiveness and efficiency of their problem-solving processes. A unique process of self-discovery lies at the core of SELF-DISCOVER, which empowers LLMs to sift through a repertoire of atomic reasoning modules, i.e., basic, fundamental components of reasoning such as critical thinking, decomposition, and step-by-step procedural thinking.

The team has shared that the LLM chooses these modules and combines them into a clear and cohesive logical structure. The LLM then follows this systematic approach in the decoding phase, directing the model through the problem-solving process in a way that more closely resembles human reasoning than ever before.

Upon evaluation, SELF-DISCOVER demonstrated a performance boost across a range of demanding reasoning benchmarks. It showed that it could improve the performance of models such as GPT-4 and PaLM 2 by up to 32% over conventional Chain of Thought (CoT) methods in tasks given by BigBench-Hard, grounded agent reasoning scenarios, and complicated mathematical problem sets (MATH). This significant performance improvement is not limited to numbers as it also signifies a significant advance in the models’ grasp and navigation of intricate issue domains.

In comparison with inference-intensive approaches like CoT-Self-Consistency, which likewise seek to improve reasoning abilities, SELF-DISCOVER has distinguished itself by its higher performance and efficiency. It surpassed these approaches by over 20% in certain instances. The team has shared that it required 10–40 times fewer inference calculations to produce these amazing outcomes despite having a far lower processing demand. This feature of SELF-DISCOVER highlights how applicable it may be in real-world scenarios, which makes it a more viable and approachable option for improving LLM reasoning skills.

In conclusion, SELF-DISCOVER is a big step forward in the search for LLMs with more complex and human-like reasoning abilities. It creates new opportunities for more effective and efficient approaches to difficult reasoning problems by empowering models to autonomously find and use task-specific reasoning structures, closing the gap between Artificial Intelligence and human cognitive processes.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.