Towards Artificial General Intelligence: Enhancing LLMs capability for Abstraction and Reasoning

Abstract

Artificial General Intelligence (AGI) represents a transformative leap in the field of artificial intelligence, aiming to achieve human-like cognitive abilities across diverse tasks. This paper explores the potential of AGI to enhance Large Language Models (LLMs) in performing abstraction and reasoning tasks, which are critical for complex problem-solving and decision making processes. Current LLMs excel in tasks involving pattern recognition, natural language understanding, and contextual generation, yet they often struggle with tasks requiring deep abstraction and logical reasoning due to their reliance on statistical correlations rather than true comprehension. We propose a novel framework that integrates AGI-driven modules with LLMs to augment their capabilities in these areas. The framework leverages AGI’s ability to model and simulate human-like thought processes, enabling the LLM to perform higher-order reasoning, draw abstract inferences, and generalize knowledge across domains more effectively. Through this integration, the enhanced LLMs demonstrate improved performance on benchmark tasks requiring complex reasoning and abstraction, such as mathematical problem-solving, scientific hypothesis generation, and advanced strategic planning. Experimental results reveal that the AGI-augmented LLMs outperform traditional LLMs in both accuracy and efficiency, particularly in scenarios that necessitate the understanding of abstract concepts and multi-step logical deductions. Additionally, this hybrid approach shows promise in reducing the need for extensive task-specific fine-tuning, thereby making LLMs more adaptable to novel problems. We utilize the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI) benchmark which measures an AI system's ability to efficiently learn new skills. We conclude that the synergistic combination of AGI and LLMs paves the way for a new generation of AI systems capable of not only processing and generating language but also engaging in sophisticated reasoning akin to human cognition.

The solution proposed in this work has won Silver Medal ( 37th Rank) in the recently concluded ARC
Prize 2024 competition.