Large Language Model and Crowdsourced Computation for AI-assisted Programming and Its Applications

Abstract

Artificial Intelligence (AI) has seen transformative advancements, significantly enhancing capabilities across various fields. This thesis explores the deep integration of large language models (LLMs) into complex problem-solving and development processes, particularly in mathematical reasoning in geometry, AI-assisted programming, and human-assisted computation. By leveraging the power of LLMs, this research demonstrates how AI enhances traditional methodologies and pioneers innovative frameworks that revolutionize understanding, generation, and reasoning with human language, source code, and visual data. These enhancements are particularly evident in developing novel AI-driven tools and systems that facilitate a more efficient and insightful interaction between humans and technology.

In the first study of this thesis, we examine a visual reasoning framework that utilizes deep learning to address constructible problems in geometry. As AI continues to advance in the perceptual systems, it increasingly emulates human sensory perception, allowing machines to interpret visual, auditory, and other sensory inputs and integrate these capabilities into complex problem-solving environments. The proposed visual reasoning framework explicitly enhances machines' capacity to engage in visual reasoning, mimicking the human ability to recognize patterns and analyze geometric features dynamically. The future integration of LLMs promises to further augment this framework by providing advanced language and geometry understanding capabilities, bridging the gap between numerical geometry processing and linguistic model comprehension, and enhancing the system's ability to interact with and make sense of geometric constructions.

The second topic extends the application of AI from geometric visual reasoning to AI-assisted programming, focusing on the profound influence of LLMs in this sector. This study highlights how integrating LLMs into software development processes complements and significantly enhances programmer capabilities through advanced code understanding and generation. This synergy between AI and programming boosts productivity and fosters the development of more innovative, efficient, and reliable software solutions. To demonstrate this integration practically, we showcase the prompting of cloud-based LLMs with the existing programming development environment to support software engineers in program composition and design. This tool exemplifies the practical application of AI insights gained from visual reasoning in geometry and programming, underlining the interconnected enhancement of AI applications across different domains.

In the final study, we explore the fusion of human-assisted computation with reinforcement learning (RL) techniques applied to LLMs for enhancing code generation, a crucial element of AI-assisted programming. This segment introduces the RL from Human Feedback (RLHF) concept, where human assessments serve as a dynamic feedback mechanism to refine RL model training. This approach marks a significant advancement within AI and effectively bridges the intricate demands of code generation with nuanced human evaluation. By integrating diverse human insights into the reward mechanisms and aligning their assessments, the framework significantly increases the accuracy and precision of LLMs in transforming textual descriptions into functional code. This method demonstrates the potential of human-AI collaboration to tackle complex programming tasks, thereby enhancing the capabilities of AI systems to understand and execute software development tasks more effectively.

This thesis demonstrates how advanced AI, particularly through the use of LLMs, is reshaping the landscape of problem-solving across various domains, from geometry to software development. By integrating LLMs with cutting-edge techniques in visual reasoning, AI-assisted programming, and human-assisted computation, we not only enhance existing problem-solving capabilities but also open new avenues for future innovations. As AI continues to evolve, the interplay between human expertise and machine intelligence will further expand, promising to unlock even more significant potential in computational methods and applications. This work lays a robust foundation for future research aimed at exploring these synergies, ultimately driving forward the boundaries of what AI can achieve across diverse fields.

Date of Award	20 May 2024
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Chung CHAN (Supervisor), Kai LIU (Co-supervisor) & Chee Wei TAN (External Co-Supervisor)