A Comprehensive Review of Large Language Models for Code Generation: Challenges and Future Directions
Authors:-Madhav Vyas, Dhruvi Dave, Khushaliba Gohil, Professor Mansi Gosai
Abstract-Large Language Models (LLMs) have significantly transformed the field of code generation by automating programming tasks, improving developer productivity, and enabling rapid prototyping. This review explores recent advancements in LLM-based code generation, examining both proprietary (closed-source) and open-source models. Proprietary models, such as GitHub Copilot, OpenAI Codex, and Amazon CodeWhisperer, offer high accuracy and seamless integration with development environments but limit user control. In contrast, open-source models like Code Llama, StarCoder, and PolyCoder provide transparency, customization, and self-hosting capabilities. Despite their progress, LLM-generated code faces challenges, including incorrect outputs, inefficiency, security risks, and difficulty in real-world software development. Benchmark datasets like HumanEval, MBPP, APPS, and CodeXGLUE have been developed to evaluate model performance based on correctness, efficiency, and robustness. Recent studies propose new methodologies, such as reinforcement learning and self-checking systems, to enhance accuracy and usability. Future research should focus on improving evaluation methods, contextual understanding, and security measures to ensure reliable and efficient LLM-generated code.
