Event logs record the execution of business processes as sequences of timestamped events. Most predictive process monitoring methods still learn a separate model per log: when the process, the activity vocabulary, or the time scale changes, the model must be retrained and revalidated. This paper takes a different route and argues for a foundation model for process mining: a single reusable model trained across heterogeneous event logs, designed to generalize to new logs and to be adapted with only a small set of in-log examples. To the best of our knowledge, we present the first event-log-native foundation model specifically tailored to process mining. The model consumes prefixes of cases directly (as ordered event sequences with timestamps and attributes), produces a compact representation of the running case, and supports two core monitoring tasks: next-activity prediction and remaining-time estimation, without any per-log parameter updates. At use time, it adapts in context from a support set sampled from the target log, which aligns the model to the local activity set and temporal scale; string attributes can optionally provide additional semantic hints. We study both balanced few-shot contexts and retrieval-based contexts built by nearest-neighbor selection with same-case masking, and we report extensive ablations over context size and design choices. Results on held-out logs and a real-life case study show that a single pretrained model can transfer to unseen event logs and improve with a handful of contextual examples. Overall, this work is a first step toward general-purpose, reusable foundation models for process mining that can be trained once and applied broadly across organizations and processes.
Berti et al. (Thu,) studied this question.