Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bias Unveiled: Investigating Social Bias in LLM-Generated Code

Lin Ling, Fazle Rabbi, Song Wang, Jinqiu Yang

TL;DR

This work addresses the problem of social bias in LLM-generated code by introducing Solar, a black-box fairness framework, and SocialBias-Bench, a dataset of 343 real-world, human-centered coding tasks. Solar automatically generates domain-specific prompts and executable test cases to quantify bias using Code Bias Score $CBS$, Bias Leaning Score $BLS$, and Pass@attribute, then leverages iterative prompting with feedback to mitigate bias while preserving functional correctness. Evaluation across four state-of-the-art code-generation LLMs reveals pervasive biases, with substantial model- and demographic-dimension variance; iterative prompting with Solar achieves substantial bias reductions (e.g., from 60.58% to 8.77% CBS in GPT-3.5-turbo-0125) and improves functional alignment. The work contributes an extensible dataset, a metamorphic-testing-inspired fairness framework, and practical bias-mitigation strategies, advancing responsible deployment of LLMs in software development.

Abstract

Large language models (LLMs) have significantly advanced the field of automated code generation. However, a notable research gap exists in evaluating social biases that may be present in the code produced by LLMs. To solve this issue, we propose a novel fairness framework, i.e., Solar, to assess and mitigate the social biases of LLM-generated code. Specifically, Solar can automatically generate test cases for quantitatively uncovering social biases of the auto-generated code by LLMs. To quantify the severity of social biases in generated code, we develop a dataset that covers a diverse set of social problems. We applied Solar and the crafted dataset to four state-of-the-art LLMs for code generation. Our evaluation reveals severe bias in the LLM-generated code from all the subject LLMs. Furthermore, we explore several prompting strategies for mitigating bias, including Chain-of-Thought (CoT) prompting, combining positive role-playing with CoT prompting and dialogue with Solar. Our experiments show that dialogue with Solar can effectively reduce social bias in LLM-generated code by up to 90%. Last, we make the code and data publicly available is highly extensible to evaluate new social problems.

Bias Unveiled: Investigating Social Bias in LLM-Generated Code

TL;DR

This work addresses the problem of social bias in LLM-generated code by introducing Solar, a black-box fairness framework, and SocialBias-Bench, a dataset of 343 real-world, human-centered coding tasks. Solar automatically generates domain-specific prompts and executable test cases to quantify bias using Code Bias Score , Bias Leaning Score , and Pass@attribute, then leverages iterative prompting with feedback to mitigate bias while preserving functional correctness. Evaluation across four state-of-the-art code-generation LLMs reveals pervasive biases, with substantial model- and demographic-dimension variance; iterative prompting with Solar achieves substantial bias reductions (e.g., from 60.58% to 8.77% CBS in GPT-3.5-turbo-0125) and improves functional alignment. The work contributes an extensible dataset, a metamorphic-testing-inspired fairness framework, and practical bias-mitigation strategies, advancing responsible deployment of LLMs in software development.

Abstract

Large language models (LLMs) have significantly advanced the field of automated code generation. However, a notable research gap exists in evaluating social biases that may be present in the code produced by LLMs. To solve this issue, we propose a novel fairness framework, i.e., Solar, to assess and mitigate the social biases of LLM-generated code. Specifically, Solar can automatically generate test cases for quantitatively uncovering social biases of the auto-generated code by LLMs. To quantify the severity of social biases in generated code, we develop a dataset that covers a diverse set of social problems. We applied Solar and the crafted dataset to four state-of-the-art LLMs for code generation. Our evaluation reveals severe bias in the LLM-generated code from all the subject LLMs. Furthermore, we explore several prompting strategies for mitigating bias, including Chain-of-Thought (CoT) prompting, combining positive role-playing with CoT prompting and dialogue with Solar. Our experiments show that dialogue with Solar can effectively reduce social bias in LLM-generated code by up to 90%. Last, we make the code and data publicly available is highly extensible to evaluate new social problems.

Paper Structure

This paper contains 13 sections, 5 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Methodology
  4. Code Bias Dataset SocialBias-Bench
  5. Generating Code Prompts
  6. Testing Code Bias
  7. Bias Mitigation Strategies
  8. Evaluation
  9. Experiment Setup
  10. Evaluation Results
  11. Results of Bias Mitigation Strategies
  12. Related Work
  13. Conclusion

Figures (4)

  • Figure 1: An overview of social bias evaluation framework Solar with examples.
  • Figure 2: An example of test case generated by Solar.
  • Figure 3: Radar chart: shape the pattern of prejudicial preferences of age on different models, the blue line: the GPT-3.5-turbo-0125, the orange line: codechat-bison@002, the green line: CodeLlama-70b-instruct-hf, the red line: claude-3-haiku-20240307. (For more information about all demographics, you can find the appendix via the shared code link.)
  • Figure 4: Illustration on the effect of hyper-parameters temperature t on CBS for the four subject LLMs. The x-axis represents the hyper-parameter values of t, while the y-axis signifies CBS.