Bias in Large Language Models: Methods, Evaluation, and Prospects

With the in-depth penetration of large language models (LLMs) such as ChatGPT and DeepSeek into critical domains including recruitment, healthcare, and finance, the issue of bias in their outputs has become a core bottleneck restricting the credible application of the technology. This paper systematically reviews the latest advances in the field of LLM bias research, classifies mainstream debiasing methods into three categories—data-level, model-level, and application-level—based on their intervention stages, elaborates on the technical logic of each category of methods, analyzes their performance in various aspects, combs through common evaluation datasets and indicator systems, and finally conducts an in-depth analysis of current research limitations and proposes targeted solutions. This paper meticulously classifies mainstream methods in recent years according to their action stages and principles, and from a practical perspective, selects factors such as interpretability, cost, and closed-source adaptability for evaluation, which are visualized as radar charts. This facilitates the analysis of the applicable scenarios of the three categories of methods, aims to analyze the advantages and disadvantages of mainstream methods, clearly presents the current research status in this field, and provides ideas for future research.