Harnessing Advanced LLMs for Fair and Efficient Performance Evaluation: An Empirical Study on DeepSeek

With the development of Large Language Model (LLM) technologies, their application potential in intelligent office systems and human resource management has become increasingly evident. To explore their applicability and effectiveness in performance evaluation scenarios, this study takes DeepSeek-V3.2 as the research subject and, based on real disciplinary development performance data from a medical college, constructs an automated evaluation framework that integrates score-based ranking and comparative (pairwise) ranking. The framework systematically examines the model’s performance in quantitative computation, qualitative analysis, ranking stability, and interpretability of generated explanations. The study uses nine anonymized performance review forms and, through prompt engineering, builds standardized input templates. Fifty repeated experimental runs are conducted under both Non-thinking Mode and Thinking Mode to test the stability and consistency of the outputs. The results show that in Non-thinking Mode, DeepSeek performs outstandingly in terms of efficiency and stability: it can rapidly complete performance analysis and ranking, and its semantic understanding helps reduce subjective bias introduced by rhetorical language. In Thinking Mode, although the model demonstrates stronger reasoning and explanatory abilities, its output variability and time cost are higher, making it more suitable for multi-round, integrative analyses. Overall, the study verifies the feasibility and application potential of LLMs such as DeepSeek in performance evaluation. In terms of data processing efficiency, objectivity, and interpretability, these models provide strong support for building intelligent performance management systems and offer empirical evidence for the development of human–machine collaborative performance evaluation models.