Temperature Effects on Prompt-Based Keyphrase Generation with Instruction-Based Large Language Models

Large language models (LLMs) demonstrate high proficiency in generating coherent and contextually appropriate texts based on given instructions. However, the influence of model parameters on task-specific performance remains underexplored. This study investigates the impact of the temperature on the stability of instruction-based LLMs in the task of generating keyphrases—one of the fundamental tasks of information retrieval—to enable the systematization of texts and speed up the search process. In experiments with three Russian-language large models (T-lite, YandexGPT, and Saiga), we performed a comparative analysis of keyphrases generated with identical prompts across different temperature settings. Our results demonstrate that increased temperature correlates with higher lexical diversity in both keyphrases and unigrams, alongside a higher incidence of keyphrases not present in the source text. The magnitude of these effects varies by model architecture. These findings underscore the importance of precision in both model selection and temperature value tuning for prompt-based keyphrase generation. The experiments were conducted using two text collections: scientific literature and news media.