Efficient Online Continual Foundation Model Fine-Tuning for Predictive Process Monitoring

Sjoerd van Straten; Marwan Hassani

cs.LGAug 28, 2026

Efficient Online Continual Foundation Model Fine-Tuning for Predictive Process Monitoring

Authors: Sjoerd van Straten, Marwan Hassani

Organizations: Department of Mathematics and Computer Science, Eindhoven University of Technology, The Netherlands · Department of Mathematics and Computer Science, University of Wuppertal, Germany

Abstract

Predictive Process Monitoring (PPM) models are increasingly deployed in dynamic environments where concept drift causes the underlying process distribution to shift over time. While recent work has moved toward online continual learning, existing methods train compact, task-specific networks entirely from scratch, leaving a persistent cold-start problem. Foundation Models (FMs) offer a compelling solution to this problem, but their continual fine-tuning in the process mining domain remains unexplored. We propose COMPASS (Continual Online foundation Model-based PPM with Adaptive SubSpaces), the first framework for online continual fine-tuning of FMs for PPM. COMPASS adapts loss-plateau drift detection to autonomously identify task boundaries in event streams and maintains a unified knowledge subspace including both pre-trained and task-specific directions. We evaluate our approach on nine event streams covering synthetic and real-world concept drift scenarios, across task-free and task-aware settings with multiple backbones and with consistent hyperparameter tuning across all methods. Our approach outperforms three SOTA non-FM competitors and two update strategy baselines, with particularly strong gains on streams exhibiting recurrent drift and complex, long-running cases, while incurring acceptable computational overhead compared to the non-FM competitors.

Figures & tables

Figure 1: Comparison of an image stream and a process event stream under concept drift. In an image stream ( left ) tasks are visually distinct between $\mathcal{T}_{A}$ and $\mathcal{T}_{B}$ . In a process event stream ( right ), cases ( $C_{1}-C_{7}$ ) are long-running and interleaved across task boundaries, and drift is caused by organizational changes causing activities (represented as letters) and the time between them to change. Resource R1 is on maternity leave after $\mathcal{T}_{1}$ , and a policy change after $\mathcal{T}_{2}$ causes newly initiated cases to be prioritized, pushing older interleaved cases to the background. No input-space signal announces the shift, making autonomous adaptation to drift substantially harder than in image streams.

Figure 2: Overview of the COMPASS architecture. Colors indicate functional stages: LoRA fine-tuning (purple), plateau detection (blue), knowledge subspace expansion (green), and adapter reinitialization (pink), described in order below.

Symbol	Meaning	Updated
$W_{p}$	Dominant input directions of pre-trained $W_{0}$	Fixed after initial SVD
$M_{t}$	Task-specific directions accumulated across tasks $1,\dots,t$	Grows at each consolidation
$M^{\prime}_{t}=[W_{p},M_{t}]$	Unified knowledge subspace used for orthogonal projection	Grows, capped at rank $R$

Table 1: Summary of the subspaces used in knowledge subspace expansion.

Task-Aware
Method	IRO5000	ORI5000	ROI5000	OIR5000	RIO5000	BPI15-REC	BPI20-RFP	BPI20-DD	BPI20-ID
DoNothing [ 13 ]	.192 $\pm$ .107	.200 $\pm$ .068	.223 $\pm$ .097	.205 $\pm$ .072	.171 $\pm$ .072	.013 $\pm$ .006	.511 $\pm$ .088	.495 $\pm$ .090	.247 $\pm$ .008
$\gamma$ = LastDrift [ 12 ]	.220 $\pm$ .085	.770 $\pm$ .003	.246 $\pm$ .056	.709 $\pm$ .003	.269 $\pm$ .052	.484 $\pm$ .000	.886 $\pm$ .001	.863 $\pm$ .001	.840 $\pm$ .006
DynaTrainCDD [ 6 ]	.775 $\pm$ .005	.785 $\pm$ .006	.790 $\pm$ .001	.729 $\pm$ .004	.784 $\pm$ .004	.455 $\pm$ .008	.831 $\pm$ .010	.830 $\pm$ .011	.793 $\pm$ .003
TFCLPM [ 14 ]	.803 $\pm$ .001	.817 $\pm$ .002	.825 $\pm$ .011	.775 $\pm$ .002	.814 $\pm$ .001	.473 $\pm$ .003	.863 $\pm$ .009	.876 $\pm$ .003	.820 $\pm$ .002
CNAPwP [ 3 ]	.802 $\pm$ .002	.816 $\pm$ .005	.829 $\pm$ .002	.780 $\pm$ .001	.812 $\pm$ .006	.492 $\pm$ .015	.879 $\pm$ .001	.881 $\pm$ .008	.846 $\pm$ .003

Table 2: Average accuracy for all baselines and datasets. Bold denotes the highest accuracy, italic and underlined the second and third highest. Sig. = statistically significant difference (

\alpha

= .05).

Figure 3: Accuracy at given index for the BPI15-REC and OIR5000 stream, illustrating COMPASS’ faster recovery after concept drift and mitigation of the cold-start problem.

Figure 4: Accuracy versus runtime (top) and memory consumption (bottom) over the event stream for ORI5000, BPI15-REC and BPI20-DD. COMPASS shows state-of-the-art accuracy with acceptable computational overhead.

$W_{p}$	$M_{t}$	TF	Avg.
$\times$	$\times$	✓	.735 $\pm$ .002
$\times$	$\times$	$\times$	.758 $\pm$ .003
✓	$\times$	✓	.721 $\pm$ .004
✓	$\times$	$\times$	.747 $\pm$ .002
✓	✓	✓	.802 $\pm$ .004 *
✓	✓	$\times$	.802 $\pm$ .003 *

Table 3: Ablation study using accuracy for our components (TF = Task-Free).

Explore similar work

Revisiting Predictive Process Monitoring in the Age of Foundation Models: A Comparative Study of Sequence, Tabular, and LLM Approaches
Jul 30, 2026Lennart Fertig, Lukas Kirchdorfer, Tobias SesterhennTabular Foundation Models Predictive Process Monitoring
University of Mannheim, Mannheim, Germany · SAP Signavio, Walldorf, Germany · Technical University of Clausthal, Clausthal, Germany
D-TAIA: Domain-Aware LLM Adaptation for Multi-Task Predictive Process Monitoring
Aug 28, 2026Sjoerd van Straten, Christine Jacob, Marwan HassaniPredictive Process Monitoring Domain Adaptation
Department of Mathematics and Computer Science, Eindhoven University of Technology, The Netherlands · Department of Mathematics and Computer Science, University of Wuppertal, Germany
LargeMonitor: Monitoring Online Task-Free Continual Learning via Large Pretrained Models
Jun 8, 2026Mingqi Yuan, Xiaoquan Sun, Shihao Luo +1Distribution Shift Concept Drift
Agentic Intelligence Lab | HKU · Qicore Tech

Jul 30, 2026J/K move · Enter open · S save

Revisiting Predictive Process Monitoring in the Age of Foundation Models: A Comparative Study of Sequence, Tabular, and LLM Approaches

Lennart Fertig, Lukas Kirchdorfer, Tobias Sesterhenn

University of Mannheim, Mannheim, Germany · SAP Signavio, Walldorf, Germany · Technical University of Clausthal, Clausthal, Germany

Tabular Foundation Models Predictive Process Monitoring Process Mining Sequence Modeling

Open paper AskPDF arXiv

Predictive process monitoring (PPM) leverages event logs to forecast the future of running process instances, for instance, predicting the next activity, the remaining time until case completion, or the time to the next event. While PPM research in recent years has been dominated by deep sequence models trained from scratch, such as Long Short-Term Memory (LSTM) models, foundation-model approaches---particularly large language models (LLMs)---are increasingly explored for PPM. At the same time, tabular foundation models with in-context learning capabilities offer a promising alternative but have not yet been systematically benchmarked for PPM. Thus, it remains unclear whether classical sequence-based models remain competitive in this evolving landscape. This paper compares the three modeling paradigms both conceptually and empirically through a controlled benchmark across multiple datasets and prediction tasks. The results show that sequence models consistently perform best for next activity prediction, whereas tabular foundation models are competitive on temporal tasks, with LLMs usually lagging behind despite higher cost.

Abstract

Explore similar work

Revisiting Predictive Process Monitoring in the Age of Foundation Models: A Comparative Study of Sequence, Tabular, and LLM Approaches

D-TAIA: Domain-Aware LLM Adaptation for Multi-Task Predictive Process Monitoring

LargeMonitor: Monitoring Online Task-Free Continual Learning via Large Pretrained Models