An Empirical Study of the Non-determinism of ChatGPT in Code Generation
There has been a recent explosion of research on Large Language Models (LLMs) for software engineering tasks, in particular code generation. However, results from LLMs can be highly unstable; nondeterministi- cally returning very different code for the same prompt. Such non-determinism affects the correctness and consistency of the generated code, undermines developers’ trust in LLMs, and yields low reproducibility in LLM-based papers. Nevertheless, there is no work investigating how serious this non-determinism threat is.
To fill this gap, this paper conducts an empirical study on the non-determinism of ChatGPT in code generation. We chose to study ChatGPT because it is already highly prevalent in the code generation research literature. We report results from a study of 829 code generation problems across three code generation benchmarks (i.e., CodeContests, APPS, and HumanEval) with three aspects of code similarities: semantic similarity, syntactic similarity, and structural similarity. Our results reveal that ChatGPT exhibits a high degree of non-determinism under the default setting: the ratio of coding tasks with zero equal test output across different requests is 75.76%, 51.00%, and 47.56% for three different code generation datasets (i.e., CodeContests, APPS, and HumanEval), respectively. In addition, we find that setting the temperature to 0 does not guarantee determinism in code generation, although it indeed brings less non-determinism than the default configuration (temperature=1). In order to put LLM-based research on firmer scientific foundations, researchers need to take into account non-determinism in drawing their conclusions.
Tue 24 JunDisplayed time zone: Amsterdam, Berlin, Bern, Rome, Stockholm, Vienna change
10:30 - 12:30 | Code Generation 2Research Papers / Journal First at Cosmos Hall Chair(s): Reyhaneh Jabbarvand University of Illinois at Urbana-Champaign | ||
10:30 20mTalk | An Empirical Study of the Non-determinism of ChatGPT in Code Generation Journal First Shuyin Ouyang King's College London, Jie M. Zhang King's College London, Mark Harman Meta Platforms, Inc. and UCL, Meng Wang University of Bristol | ||
10:50 20mTalk | Don’t Complete It! Preventing Unhelpful Code Completion for Productive and Sustainable Neural Code Completion Systems Journal First Zhensu Sun Singapore Management University, Xiaoning Du Monash University, Fu Song Institute of Software at Chinese Academy of Sciences; University of Chinese Academy of Sciences; Nanjing Institute of Software Technology, Shangwen Wang National University of Defense Technology, Mingze Ni University of Technology Sydney, Li Li Beihang University, David Lo Singapore Management University | ||
11:10 20mTalk | Divide-and-Conquer: Generating UI Code from Screenshots Research Papers Yuxuan Wan The Chinese University of Hong Kong, Chaozheng Wang The Chinese University of Hong Kong, Yi Dong The Chinese University of Hong Kong, Wenxuan Wang Chinese University of Hong Kong, Shuqing Li The Chinese University of Hong Kong, Yintong Huo Singapore Management University, Michael Lyu Chinese University of Hong Kong DOI | ||
11:30 20mTalk | LLM-based Method Name Suggestion with Automatically Generated Context-Rich Prompts Research Papers Waseem Akram Beijing Institute of Technology, Yanjie Jiang Peking University, Yuxia Zhang Beijing Institute of Technology, Haris Ali Khan Beijing Institute of Technology, Hui Liu Beijing Institute of Technology DOI | ||
11:50 20mTalk | Beyond Functional Correctness: Investigating Coding Style Inconsistencies in Large Language Models Research Papers Yanlin Wang Sun Yat-sen University, Tianyue Jiang Sun Yat-sen University, Mingwei Liu Sun Yat-Sen University, Jiachi Chen Sun Yat-sen University, Mingzhi Mao Sun Yat-sen University, Xilin Liu Huawei Cloud, Yuchi Ma Huawei Cloud Computing Technologies, Zibin Zheng Sun Yat-sen University DOI | ||
12:10 20mTalk | Refining ChatGPT-Generated Code: Characterizing and Mitigating Code Quality Issues Journal First Yue Liu Monash University, Le-Cong Thanh The University of Melbourne, Ratnadira Widyasari Singapore Management University, Singapore, Kla Tantithamthavorn Monash University, Li Li Beihang University, Xuan-Bach D. Le University of Melbourne, David Lo Singapore Management University | ||
This is the main event hall of Clarion Hotel, which will be used to host keynote talks and other plenary sessions. The FSE and ISSTA banquets will also happen in this room.
The room is just in front of the registration desk, on the other side of the main conference area. The large doors with numbers “1” and “2” provide access to the Cosmos Hall.