PM-LLM-Benchmark: Evaluating Large Language Models on Process Mining Tasks

TL;DR

PM-LLM-Benchmark addresses the lack of PM-specific benchmarks for evaluating LLMs by proposing a modular benchmark that tests domain knowledge and implementation strategies across two modes: direct insights and code generation. The evaluation relies on an LLM judge rather than ground-truth answers to handle open-ended PM queries, acknowledging biases and limitations. Empirical results suggest that large commercial and open-source LLMs approach PM task adequacy, while tiny edge devices remain insufficient, and aggressive quantization harms performance. The work provides a foundation for comparing LLMs in PM, highlights biases in evaluation, and outlines directions such as Retrieval-Augmented Generation, agent-based benchmarking, and dynamic data generation to improve realism and ranking.

Abstract

Large Language Models (LLMs) have the potential to semi-automate some process mining (PM) analyses. While commercial models are already adequate for many analytics tasks, the competitive level of open-source LLMs in PM tasks is unknown. In this paper, we propose PM-LLM-Benchmark, the first comprehensive benchmark for PM focusing on domain knowledge (process-mining-specific and process-specific) and on different implementation strategies. We focus also on the challenges in creating such a benchmark, related to the public availability of the data and on evaluation biases by the LLMs. Overall, we observe that most of the considered LLMs can perform some process mining tasks at a satisfactory level, but tiny models that would run on edge devices are still inadequate. We also conclude that while the proposed benchmark is useful for identifying LLMs that are adequate for process mining tasks, further research is needed to overcome the evaluation biases and perform a more thorough ranking of the competitive LLMs.