Exploring Variable Potential for LLM-based Log Parsing Efficiency and Reduced Costs

Log parsing extracts structured events from massive system logs and is essential for tasks like compression, anomaly detection, and failure diagnosis. With the rise of LLMs, their strong text understanding and summarization capabilities have enabled more accurate log parsing. However, existing LLM-based methods mainly focus on log constants, overlooking the value of log variables, leading to inefficient sampling, cache use, and context learning. To address this, we propose a variable-centric strategy called VISTA, which fully leverages log variables through dynamic contribution sampling, variable-centric cache, and adaptive variable-aware ICL. Early results show improved parsing efficiency and significantly reduced LLM usage costs.