An Empirical Study of Exploring the Capabilities of Large Language Models in Code Learning

Shangqing Liu, Daya Guo, Jian Zhang, Wei Ma, Yanzhou Li, Yang Liu

Abstract: Since the advent of ChatGPT, large language models (LLMs) have attracted widespread attention from academia and industry. They have also brought significant changes to software engineering. However, until now, there has been a lack of comprehensive studies comparing LLMs with previous smaller code pre-trained models. To address this gap, we conduct a study in this paper to illustrate the performance of LLMs in different software engineering tasks. Specifically, we select three open-source large language models, CodeGen, LLaMA, and StarCoder, for the research targets, and our study is conducted from four aspects, including code syntax understanding, code semantic reasoning, encoding representation quality, and adaptation performance for different software engineering tasks to compare LLMs with previous code pre-trained models. Four aspects build on each other, forming important components of AI for Software Engineering. We conclude that: (1) Compared with previous smaller pre-trained models like CodeBERT, LLMs exhibit distinct trends in how they learn code syntax or semantics as the number of layers increases. Additionally, mastering code semantics proves to be more challenging, with semantic information usually learned in the final layers; (2) Causal decoder architecture with left-to-right attention masking does not perform well in zero-shot tasks; (3) For classification tasks, the mean vector representation generated by LLMs over a sequence tends to outperform the last token representation in the sequence; (4) Incorporating parameter-efficient fine-tuning techniques into LLMs for downstream tasks can help LLMs achieve better performance than previous code pre-trained models on code generation tasks but may not be optimal in some code understanding tasks; (5) LoRA emerges as a more effective PEFT technique for LLMs in downstream code-related tasks. We hope these findings will better guide future researchers in designing more powerful code models.

External IDs:dblp:journals/tse/LiuGZMLL25