A machine-learning framework called COMPASS recorded the highest next-activity prediction accuracy on seven of nine event streams in a benchmark of changing business processes, according to an arXiv preprint. On the real-world BPI15-REC stream, the paper reports a 19% gain over the competing baseline CNAPwP.
The work tackles online predictive process monitoring, where a model predicts what activity will come next while a stream of events changes. The authors propose COMPASS for continual fine-tuning of a foundation model, meaning a pretrained model adapted repeatedly as new process data arrive.
How COMPASS protects earlier knowledge
COMPASS consolidates pretrained and task-specific directions into a unified knowledge space. It then reinitializes new adapters in the orthogonal residual space, meaning a separate direction outside the information already stored. The design separates new adapter updates from the consolidated directions.
In task-free operation, consolidation is triggered by the variation in recent window losses, or prediction errors. When that rolling variation falls below a preset threshold, the system identifies a stable loss plateau and triggers consolidation. A cooldown period is used to prevent repeated triggers.
Seven wins across nine streams
The test covered nine event streams: five synthetic streams and four real-world BPI streams. The researchers reported full-stream average accuracy for next-activity prediction, per-window accuracy, runtime and memory consumption.
For task-free COMPASS using DistilGPT2, the displayed mean accuracies ran from .676 to .893 across the dataset columns. The figures were reported as means and standard deviations over five runs with different random seeds. Performance therefore varied across the streams, even though COMPASS led on most of them.
The task-free version also closely tracked the task-aware oracle across all datasets. That comparison suggests that, in these tests, the automatic consolidation rule came close to the reference configuration that had task information available.
The design’s moving parts
An ablation of the model’s stored knowledge components reported .802 accuracy when both the pretrained subspace, Wp, and the task-specific subspace, Mt, were used. Removing both produced .735 in the task-free setting and .758 in the task-aware setting. Keeping Wp alone produced .721 and .747, respectively.
COMPASS was compared with five baselines: two update strategies and three non-foundation-model competitors. It was tested with Tiny-LLM, which has 10 million parameters, and DistilGPT2, which has 82 million. The experiments ran on a Tesla V100 with 16GB of GPU memory, and each result was averaged over five repetitions with different seeds.
The online evaluation used a test-then-train arrangement. The first 15% of cases served for validation, while the remaining 85% were used for online testing followed by training updates. The main comparison used a paired two-sided Student’s t-test, with a significance threshold of 0.05, on matched per-seed accuracies.
Accuracy came with a time cost
The accuracy lead came with a computing cost. COMPASS with DistilGPT2 was slower than lightweight baselines. The γ = LastDrift strategy had comparable or higher runtime on several datasets, but its accuracy was consistently lower in the reported comparison. The supplied analysis does not give exact runtime figures.
The report says memory remained stable after the one-time extraction of Wp, with a rank cap preventing unbounded growth of the stored subspaces. Exact memory values were not provided in the supplied analysis.
What the benchmark leaves open
These are benchmark results rather than a direct test of operational or business outcomes. The evaluation covered nine event streams, two model backbones and the listed accuracy, runtime and memory endpoints. That makes the comparison informative for the tested configurations, but leaves open how COMPASS would perform on other streams, models, hardware or deployment conditions.
The study focused on next-activity prediction. It therefore does not answer whether the same approach would work for final-outcome prediction, remaining-time prediction or other predictive-process-monitoring tasks. The reported results support comparative accuracy claims on the tested streams, not a general guarantee beyond them.
The document is identified as arXiv:2608.28237v1 and dated 28 August 2026. It is a preprint.
Paper data and sources
Original title: Efficient Online Continual Foundation Model Fine-Tuning for Predictive Process Monitoring
Authors: Sjoerd van Straten, Marwan Hassani
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text