Author-updated Summary
Verified author edit
Predictive Process Monitoring (PPM) aims to forecast future process outcomes—such as the next activity, remaining time, or time until the next event—using historical event logs. While deep sequence models (like LSTMs) have long been the standard, the rise of foundation models has introduced new alternatives, including Large Language Models (LLMs) and tabular foundation models. This paper provides a systematic, controlled benchmark comparing these three paradigms to determine if classical sequence models remain the most effective approach.
The authors categorize PPM approaches into three distinct paradigms:
The empirical benchmark across five real-world event logs reveals that no single paradigm dominates every task. Sequence models (LSTM and Transformer) consistently achieve the highest accuracy for next-activity prediction, likely because they explicitly model the sequential trajectory of events. In contrast, tabular foundation models are highly competitive for temporal tasks (remaining time and next-event timing), often matching or exceeding sequence models. LLMs generally underperform compared to the other two paradigms while incurring significantly higher computational costs.
Alex: Welcome to another episode of ResearchPod.
Sam: Today we're looking at a question that's become increasingly pressing in the world of data science. Businesses run on processes—supply chains, hospital admissions, loan approvals—and for years, researchers have built specialized software to predict what happens next in those processes. Now that large, general-purpose AI models are everywhere, the question is: should we replace those specialized tools with the new ones?
Alex: And I'm guessing the answer isn't a simple yes or no.
Sam: Not at all. The paper we're discussing, "Revisiting Predictive Process Monitoring in the Age of Foundation Models," puts three different types of models head-to-head on real business process data. The central puzzle is whether these newer, larger models actually outperform the older, purpose-built ones—or whether specialization still wins.
Alex: So what are the three types they're comparing?
Sam: Think of it like three different ways to learn a job. The first type is called a sequence model. Imagine a factory worker who has spent years on the same assembly line. They know exactly which step comes after which, because they've done it thousands of times. These models are trained from scratch on a specific process, and they develop a deep instinct for the order of events.
Alex: So they learn one process really well, but only that process.
Sam: Right. The second type is a Large Language Model—an LLM. Think of this as a very well-read assistant who has absorbed millions of documents. You give them a small amount of extra training on your specific process, and they use their broad general knowledge to guess what comes next. They haven't lived on your assembly line, but they've read about a lot of assembly lines.
Alex: And the third type?
Sam: These are called tabular foundation models. Rather than following the sequence of events like a story, they take a snapshot of the current moment—what step are we on, how much time has passed, what resources are involved—and use that snapshot to make a prediction. Where sequence models read the whole story of a process, tabular models read the current chapter summary.
This study clarifies the trade-offs between specialized sequence modeling and general-purpose foundation models in process mining. For practitioners, the results suggest that sequence models remain the robust choice for activity-based predictions, while tabular foundation models offer a viable, efficient alternative for temporal forecasting. The findings also highlight the limitations of applying LLMs to structured process data, suggesting that their general-purpose capabilities do not automatically translate to superior performance in process-specific tasks.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.
Alex: So the researchers tested all three on real data. What were they actually trying to predict?
Sam: Three things: what the next step in a process will be, how long until that next step happens, and how long until the entire process finishes. These are called Predictive Process Monitoring tasks. The idea is that if you can spot a delay early, you can act before it becomes a real problem—an early warning system for operational bottlenecks.
Alex: And what did they find?
Sam: For predicting the next activity—the actual next step—the specialized sequence models were the clear winners. The reason gets at something fundamental about how these models are built. Sequence models have a built-in instinct for order. They're wired to ask, "given everything that happened before, what comes next?" When a process has many possible branching paths—after step A, you might go to B, C, or D—that instinct is very valuable. The sequence model has learned which branches are likely from thousands of examples. The LLM is making a more educated guess based on general knowledge, and the tabular model isn't really tracking the sequence at all.
Alex: So the more complex the branching, the more the sequence model pulls ahead?
Sam: That's what the data suggests, and it held across the different event logs they tested. But here's where it gets more nuanced: for timing predictions, the picture flips.
Alex: How so?
Sam: For predicting how long a process will take, the tabular foundation models were genuinely competitive—in some cases, better than the sequence models. The reason is almost the opposite of why sequence models win on next-step prediction. For timing, you don't need to know the exact path. You need to recognize patterns in the current state of things. Tabular models were pre-trained on enormous amounts of structured data, so they've already absorbed a lot of those statistical patterns. They don't need to relearn them from scratch.
Alex: So the tabular models are better at reading the room, while the sequence models are better at following the plot.
Sam: That's a useful way to put it. And the practical implication the authors raise is that a hybrid approach might be the most sensible design for future systems—use a sequence model to navigate the path, and a tabular model to estimate the timing.
Alex: Where did the LLMs land?
Sam: Somewhere in the middle, and with a specific weakness worth noting. They generally outperformed the tabular models on next-step prediction, but they were slower and more expensive to run. They also showed a consistent failure pattern: on several datasets, they would frequently predict that a process had ended when it actually hadn't—calling the finish line too early, much more often than the sequence models did.
Alex: So it's not that they're unreliable in a random way. They have a particular blind spot.
Sam: Correct. And that kind of systematic error matters a lot in operational settings, where a false "all done" signal could cause real problems downstream.
Alex: Are there limitations to the study we should keep in mind?
Sam: The authors are quite open about a few. The LLMs they tested were relatively small versions—it's possible that much larger models might perform better, though they would also be considerably more expensive to run, which is itself a practical constraint. The tabular models also struggled with the largest datasets because they can only process a limited amount of data at once. That's a common ceiling in current AI systems, and it matters when your event logs are very long.
Alex: So the findings are meaningful, but they're a snapshot of where these models are right now—not a permanent verdict.
Sam: That's the right framing. What the study does establish clearly is that specialized sequence models remain highly effective for structured prediction tasks. The most complex model isn't automatically the most useful one. A model built specifically to understand the structure of your data can outperform a general-purpose system, even a much larger one.
Alex: It's a useful reminder that in data science, picking the right tool for the specific job still matters more than picking the most powerful tool on the shelf.
Sam: Thanks for listening to ResearchPod.