D-TAIA Enhances LLM Adaptation for Process Monitoring

Predictive Process Monitoring (PPM) is crucial for organizations to forecast future process behaviors, such as the next activity or remaining time for ongoing cases. However, existing PPM methods often struggle with real-world challenges like data scarcity, high process entropy, and distributional shifts in event logs. While Large Language Models (LLMs) offer powerful sequential reasoning capabilities, adapting them effectively to multi-task PPM under these conditions remains an open challenge. Current FM-based approaches often lack mechanisms for handling distributional shifts or rely on direct regression heads that are not well-suited for continuous time prediction. This paper introduces D-TAIA (Domain-aware Training and Attention-based Inference Architecture), a novel framework designed for joint next activity and remaining time prediction using parameter-efficient fine-tuning of an FM backbone. D-TAIA integrates domain-aware triplet loss (DATL) pre-training with FAISS-based nearest neighbor retrieval for remaining time prediction. It also employs the TAIA inference strategy to preserve pre-trained sequential reasoning during fine-tuning. Evaluated across four real-world event logs, D-TAIA consistently achieves state-of-the-art or competitive performance compared to fine-tuned LLMs and recurrent neural network baselines, demonstrating effective transfer of NLP and computer vision techniques to PPM with a compact 10M-parameter backbone.