Achieving AI-style reasoning without an LLM

Researchers at the Allen Institute for Artificial Intelligence (AI2) have made a significant breakthrough in achieving AI-style reasoning without relying on large language models (LLMs). Their approach, called Minotaur, uses a combination of modular reasoning and neural-symbolic integration to reason about complex tasks. In a paper presented at the NeurIPS conference, the researchers demonstrated that Minotaur can perform tasks such as logical reasoning, arithmetic, and planning without the need for LLMs. The system is based on a modular architecture that consists of a neural network and a symbolic reasoning module. The neural network is used to learn abstract representations of the task, while the symbolic module provides logical reasoning capabilities. The researchers achieved this by using a combination of techniques such as modular reasoning, neural-symbolic integration, and knowledge graph-based reasoning. Their approach is a significant step towards achieving AI-style reasoning without the need for LLMs, which are often criticized for their lack of transparency and interpretability. The Minotaur system is also designed to be more explainable and transparent than LLMs, as it provides clear and explicit explanations for its reasoning process. The researchers believe that Minotaur has the potential to be used in a wide range of applications, including decision-making systems, planning, and natural language processing.