The 2025 publications of the Process and Data Science (PADS) group show a field that is moving rapidly beyond classical process discovery. Looking at a selection of key papers, four developments stand out: the increasingly close relationship between process mining and Artificial Intelligence, the maturation of Object-Centric Process Mining, richer techniques for conformance checking and concurrency, and a growing emphasis on statistically sound comparison, privacy-preserving analysis, and simulation.
Generative AI continued to play an important role in 2025, but the focus is shifting from simply applying Large Language Models to process mining toward understanding where LLMs add value and where specialized process-mining techniques remain indispensable.
The paper An LLM-Based Q&A Natural Language Interface to Process Mining investigates how LLMs can make sophisticated process-mining functionality accessible through natural-language questions. Rather than asking an LLM to perform process analysis itself, the approach combines the flexibility of language models with established process-mining algorithms. Similarly, the PM-LLM-Benchmark systematically evaluates the ability of LLMs to perform process-mining tasks, showing why dedicated benchmarks are needed instead of assuming that general-purpose AI also understands processes.
At the same time, Process Mindlessness: When we Lose Sight of What AI is Supposed to Improve takes a broader perspective. AI may optimize individual tasks while making the overall process worse through bloating, blurring, and blasting. The relevant question is therefore not simply whether AI can automate a task, but whether it improves the end-to-end process.
Together, these papers reinforce an increasingly important message: AI needs process intelligence to understand the operational context in which individual tasks are embedded.
A second major trend is the continued maturation of Object-Centric Process Mining (OCPM). The focus is no longer only on representing event data without forcing everything into a single case notion. The 2025 work develops an increasingly complete toolbox for discovering, querying, and analyzing the interactions between multiple objects.
Object-Centric Local Process Models combines object-centric process mining with behavioral pattern mining, allowing interesting local behavior to be discovered without reducing a complex process to one end-to-end case notion. Object-Centric Causal Nets introduces a new modeling formalism aimed at representing complex interactions between objects without some of the representational restrictions of existing object-centric models.
This is complemented by a declarative perspective. OC-DECLARE: Discovering Object-Centric Declarative Patterns with Synchronization captures constraints involving multiple object types and their synchronization. For example, it becomes possible to express requirements concerning how an order and all of its items should interact rather than considering each object type in isolation.
Querying is another essential building block. OCPQ: Object-Centric Process Querying and Constraints introduces an expressive approach for querying object-centric event data and checking constraints. The accompanying OCPQ Tool demonstrates how such complex queries can be constructed visually and executed efficiently.
These developments show that OCPM is evolving from a new event-data perspective into a comprehensive framework for process intelligence, covering modeling, discovery, querying, constraints, and behavioral analysis.
A third recurring theme is that real event data should not automatically be interpreted as a simple sequence. Timestamps may overlap, have limited precision, or be missing altogether, and many processes contain genuine concurrency. Several 2025 papers therefore move away from the traditional assumption that all events can be placed in one total order.
Partially Ordered Stochastic Conformance Checking combines partial-order representations with stochastic conformance checking, allowing both concurrency and probabilistic behavior to be taken into account. Computing Alignments for Partially-Ordered Traces Through Petri Net Unfoldings uses Petri-net unfolding techniques to compute alignments while preserving concurrency and uncertainty in event data.
At the same time, improving the computational foundations of conformance checking remains important. A Dynamic Programming Approach for Alignments on Process Trees exploits the structure of process trees to compute optimal alignments efficiently.
The common direction is clear: process-mining algorithms are becoming more faithful to the true semantics of concurrent processes instead of imposing artificial sequences simply because traditional event logs require them.
Processes frequently extend beyond a single organization, but the corresponding data cannot always be pooled because of privacy, confidentiality, or commercial sensitivity. This motivates another important research direction: Federated Process Mining.
Federated Conformance Checking shows how conformance problems in cross-organizational processes can be analyzed while keeping sensitive local information protected. Your Secret Is Safe With Me: Federated Directly-Follows Graph Discovery addresses the complementary discovery problem and uses homomorphic encryption to construct a process view spanning organizations without revealing their underlying event data.
This is particularly relevant for processes such as supply chains and healthcare pathways, where understanding the end-to-end process requires information from multiple independent parties. Federated approaches aim to obtain system-level process intelligence without first requiring everybody to surrender their data.
Finally, process mining is increasingly moving beyond describing what happened. Two selected papers illustrate a transition toward more rigorous process comparison and forward-looking analysis.
Hypothesis Testing for Processes introduces statistical hypothesis testing for comparing processes across multiple dimensions. Rather than merely visualizing differences between two process populations, the objective is to determine whether the observed differences are statistically significant.
Reliable and Configurable Process Simulations via Probabilistic White-Box Models looks forward rather than backward. Process simulation provides a virtual environment in which alternative scenarios can be explored before changing the real process. By using probabilistic white-box models, the approach combines realistic variability with models whose decision logic remains understandable and configurable.
These developments illustrate an important progression: from discovering processes, to diagnosing and comparing them, and ultimately to exploring what will happen when we intervene.
Taken together, the selected 2025 publications show process mining entering a new phase. Object-centric techniques provide a more faithful representation of operational reality. Partial orders capture concurrency without inventing artificial sequences. Federated approaches allow analysis across organizational boundaries. Statistical methods make process comparisons more rigorous, while simulation supports experimentation before interventions are implemented. At the same time, generative AI creates new interfaces and possibilities, but also makes trustworthy operational context more important than ever.
The connecting concept is Process Intelligence: transforming event data into reliable knowledge about how processes actually operate and using this knowledge to support humans and AI systems in making better decisions.
The direction emerging in 2025 is therefore clear: from process mining as an analytical technology toward process intelligence as the operational foundation for trustworthy AI and intelligent action.
See the complete overview of the 2025 publications.