acceptodds
Under review as a conference paper at ICLR 2027

Foundation Models for Event Prediction Across Heterogeneous Event Logs

Abstract

Event logs are ubiquitous across real-world domains, including business processes, manufacturing workflows, healthcare, and many other operational systems. Yet event logs can differ substantially in their event semantics, label spaces, interaction patterns, and temporal scales, making it difficult to build a single foundation model that generalizes directly to previously unseen systems. We address this challenge with the Foundation Event Predictor (FEP), a simple and scalable in-context architecture that performs next-event prediction through learned, event-level nonparametric inference. Causal self-attention encodes context and target histories into key and query representations, respectively, while the observed next context events are encoded as values. Cross-attention then performs learned, relevance-weighted aggregation of next-event outcomes over all context transitions, allowing the entire nonparametric prediction procedure to be expressed as standard attention operations and efficiently parallelized with modern kernels such as FlashAttention. To promote cross-system generalization, we introduce permutation-robust embedding learning and random fixed embeddings to decouple event IDs from persistent semantics, together with empirical quantile transforms and robust normalization to handle diverse temporal scales. We pretrain FEP on 696 real-world event datasets spanning heterogeneous event systems and evaluate it on 50 held-out datasets unseen during pretraining. Across these datasets, FEP generalizes effectively to new event systems without parameter adaptation, frequently approaching the performance of strong models trained separately on each target dataset while substantially outperforming existing foundation-model baselines, particularly on challenging prediction tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
event modeling, event prediction, foundational models, event logs, process events, process mining, process prediction, process monitoring, heterogeneous event logs

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_uHAGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Foundation Event Predictor (FEP), an in-context foundation model for next-event prediction, including both the next event type and the inter-arrival time. It aims to generalize the foundation model to previously unseen, heterogeneous event logs without parameter adaptation. The model decomposes context sequences into event-level history-to-next-event pairs, encodes context histories as keys and target histories as queries via a shared causal Transformer, encodes observed next events as values, and performs learned nonparametric aggregation via cross-attention. FEP is pretrained on 696 real-world event datasets and evaluated on 50 held-out datasets, comparing against per-dataset local models (Transformer, GRU, Neural Hawkes, etc.) and existing foundation models (FM-v2, FIM-PP). Results show FEP outperforms several foundation models and often approaches the best local model.

Critical strengths
  1. It is an important problem to establish a foundation model generalizing across heterogeneous event logs.
  2. Pretraining on 696 datasets and testing on 50 held-out real logs is a substantial effort. The experiments are convincing.
Critical weaknesses
  1. It lacks a clear figure that conveys the model architecture and data flow. The text alone is insufficient to reproduce the model.
  2. The paper lacks a careful and complete introduction to data curation and the characteristics under different evaluation settings.
  3. More discussion and explanation are needed to analyze the contributions of the proposed model compared with the baselines and variants. For example, why does FEP-Local lag so far behind Transformer/GRU? Model size ablations show that larger size is worse, but no discussion is given.
Questions for authors
  1. Could you provide a complete architecture figure to clearly illustrate the data flow in the whole model and the model's architecture?
  2. Please detail the full preprocessing pipeline for pretraining and testing.
3 comments

Weakness 1 and Question 1: Missing architecture and data-flow figure

R1: We respectfully clarify that Figure 1 (page 4) of the main paper already provides an explicit visualization of FEP's architecture and complete data flow. This figure illustrates the core methodological contribution of our work: how context and target event sequences are encoded through shared causal self-attention, how context histories and their subsequent events form keys and values, how target histories form queries, and how cross-attention aggregates context information for next-event prediction. The figure is prominently presented in Section 3.2 and directly corresponds to the detailed mathematical formulation that follows. Furthermore, Sections 3.2–3.5 provide the mathematical definitions and model components, while Algorithm 1 in Appendix A explicitly outlines the complete pretraining pipeline, including context and query sampling, causal encoding, cross-attention, loss computation, gradient accumulation, and parameter updates. Appendix A also provides detailed implementation and training configurations. Thus, the requested architecture visualization, data flow, and pretraining pipeline are already explicitly presented and documented in the submitted paper.

Weakness 2 and Question 2: Missing dataset curation, characteristics, and preprocessing details

R2: We respectfully clarify that the submitted paper already provides extensive documentation of dataset curation, preprocessing, dataset characteristics, and experimental procedures, both in the main text and appendices. Specifically:

(1) Dataset curation, preprocessing, and characteristics (Appendix C). Appendix C explicitly describes the dataset sources, curation and preprocessing procedures, including activity-type handling and verification of non-overlap between pretraining and test datasets. Table 9 provides domain-wise statistics for all pretraining and test datasets, including the number of cases, activity types, average trace lengths, and durations. Figure 5 further visualizes the distributions of inter-arrival times, trace durations, activity counts, and trace lengths, comparing the pretraining datasets with easy, moderate, and hard test groups. Furthermore, Table 10 provides a detailed catalog of every test dataset, including their sources, descriptions, statistics, and applicable preprocessing details.

(2) Pretraining and evaluation procedures (Section 4.1 and Appendix A). Section 4.1 explicitly describes the train/validation/test splitting strategy, baseline selection and optimization, evaluation metrics, and dataset difficulty categorization. Appendix A provides further details on model configurations, training settings, baseline implementations, hyperparameter search, and validation procedures, as explicitly referenced in Section 4.1 (Line 294). Moreover, Algorithm 1 in Appendix A presents the complete episodic pretraining procedure, including context and query sampling, causal encoding, cross-attention, gradient accumulation, and parameter updates.

Thus, the requested dataset characteristics, preprocessing methodology, and pretraining/evaluation protocols are already extensively documented in the submitted paper. We believe these descriptions provide a clear basis for understanding and independently implementing the proposed methodology and evaluation procedures.

Weakness 3: Insufficient analysis of model contributions, FEP-Local, and model-size ablations

R3: We respectfully disagree that the paper lacks sufficient analysis of model contributions and behavior. Section 4.2 and Appendix B already provide extensive ablation studies, accompanied by detailed analyses and discussions of their findings, addressing several key factors affecting FEP's performance.

(1) Extensive ablation studies and analyses. We systematically investigate context size (Figure 3a), number of pretraining datasets (Table 3a), model dimension and depth (Figure 3b and Table 3b), event-ID representation strategies (Table 6), and time normalization strategies (Table 7). These experiments provide important insights into FEP's behavior, including the benefits of larger context sets and diverse pretraining datasets, the importance of removing persistent event-ID semantics, and the effectiveness of different time normalization strategies. Furthermore, Figure 4 provides dataset-wise comparisons between FEP and FEP-Local, while Appendix B analyzes inference efficiency and scalability (Tables 4–5). These results and their implications are explicitly discussed in the paper.

(2) Why FEP-Local underperforms Transformer/GRU. FEP-Local uses exactly the same architecture, hyperparameters, and training procedure as FEP; the only difference is that it is trained on a single target dataset rather than across heterogeneous datasets. No dataset-specific architecture or hyperparameter optimization is performed for FEP-Local. This controlled comparison is specifically designed to isolate the benefits of cross-dataset pretraining, rather than to establish the best achievable performance of FEP under dataset-specific optimization. In contrast, Transformer and GRU are trained and tuned separately for each dataset. Therefore, the performance gap between FEP-Local and these optimized baselines should not be interpreted as evidence of an inherently weaker architecture. The substantial improvement from FEP-Local to pretrained FEP directly demonstrates the benefits of cross-dataset pretraining, as further supported by the dataset-wise analysis in Figure 4.

(3) Clarification of model-size ablations. The reviewer's statement that larger models perform worse does not accurately reflect our results. Figure 3b shows that increasing model dimension initially yields substantial improvements in both accuracy and NMAE, with performance largely saturating around dimensions 64–128 and deteriorating only when increased further to 256. Similarly, Table 3b shows that increasing model depth from one to a few layers provides modest improvements, while excessive depth (eight layers) degrades performance. These trends are already explicitly discussed in Section 4.2 (under "Model Size"). The results indicate that FEP benefits from sufficient model capacity but does not require an excessively large backbone, with model width having a more pronounced effect than moderate increases in depth.

Thus, the submitted paper already provides systematic ablation studies and substantive analyses addressing the specific issues raised by the reviewer, including the role of cross-dataset pretraining and the non-monotonic effects of model size.

Thank you for the responses. I have read the rebuttal and have no further questions.

Reviewer_Yj2wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Foundation Event Predictor (FEP), a model that predicts the next event and its timing in event logs, from systems that it has never seen, without the need for retraining. Given example traces from a new system as context, FEP uses attention to weigh past situations that are similar to the current one. To work across unrelated systems, it avoids attaching fixed meanings to event IDs and normalizes time scales from the context. FEP is pretrained on 696 public event logs and tested on 50 held-out ones, showing that it performs better than some existing foundation models for event data and close to but not generally better than models that are trained separately on each log.

Critical strengths

A major strength of the work is that it addresses a valuable transfer problem and is backed by a substantial data effort. Event logs are ubiquitous so a single pretrained model that transfers to unseen systems, despite their unrelated labels and different time scales, would lower the barrier to prediction in many practical settings. The authors’ collection of numerous public event logs, with a held-out test set checked for overlap with the pretraining data, is a useful resource. Note that transfer in event settings has been studied in prior work.

Another strength is that the design seems simple and well-motivated overall. The core idea is intuitive: to predict what happens next, look at similar past situations in the context and combine what followed them. Implementing this with standard attention keeps the architecture simple and efficient.

Critical weaknesses

A major weakness is that empirical results do not fully make the practical case for the work. While FEP beats other foundation models, it generally does not match models trained directly on each log; its claimed benefit from pretraining is measured against a weak local version of its own architecture. In my view, the paper does not sufficiently quantify the appropriate savings or identify settings where it would matter.

Another weakness is that important baselines seem to be missing. FEP essentially predicts the next event by looking up similar situations in the context, yet it is never compared with simple methods that do the same, such as a nearest-neighbor predictor fit on the context traces. It is hard to tell how much FEP learns beyond basic transition statistics. A baseline where a pretrained model is fine-tuned would also show whether light adaptation outperforms FEP’s no-update approach (see my next comment).

The paper focuses heavily on process mining and seems to be missing some key references. For instance, it omits the substantial line of work on pretraining for event sequences, which is the most direct precedent for the paper’s central claim about transfer. Self-supervised methods such as contrastive pretraining with masked and inserted “void” events and contrastive sequence embeddings pretrain event-sequence models for transfer to downstream data, and LLM-based approaches bring pretrained knowledge to event prediction. These works differ from FEP in requiring fine-tuning or in not targeting unrelated event vocabularies, but they should be discussed. Also, a pretrain-then-fine-tune baseline in this style would help show whether FEP’s no-update, in-context approach is competitive with lightweight adaptation.

I am open to revising my score based on the discussion.

Questions for authors

Can the authors comment about the relation to other in-context and meta learning approaches? They seem highly relevant. I recommend citing closely related work on in-context learning, such as attentive neural processes and TabPFN.

What are the training times and costs of the best local models relative to FEP inference, and in which settings does FEP outperform them?

8 comments

Weakness 1: Practical value relative to locally trained models

R1: We respectfully disagree that FEP's slightly lower performance than the best locally trained models undermines its practical value or research significance.

(1) Practical value: Near-local performance without dataset-specific training. Across 50 unseen datasets, FEP achieves an average event-type accuracy only ~2 percentage points below the strongest local baseline, with a difference of merely ~0.003 in time NMAE (Table 1). Such small gaps must be weighed against the substantial practical advantages of foundation models. Unlike local models, which require separate training, architecture and hyperparameter search, validation-based model selection, and dedicated ML expertise and infrastructure for each new system, FEP performs zero-shot inference without any parameter updates. This substantially reduces the effort and complexity of deploying predictive models across heterogeneous event systems, particularly when ML resources are limited. Thus, FEP's practical value lies not in outperforming individually optimized local models at all costs, but in achieving nearly equivalent performance while avoiding extensive dataset-specific training, tuning, and model selection.

(2) Research significance: Substantial progress toward general-purpose event prediction. Requiring a zero-shot foundation model to outperform the best individually tuned local models is an overly restrictive criterion for evaluating progress in this emerging research direction. FEP already substantially outperforms existing event foundation models, including FM-v2 (Berti and van der Aalst, 2026b) and FIM-PP (Berghaus et al., ICLR 2026). For example, on hard datasets, FEP achieves 0.35 accuracy, compared with 0.24 for FM-v2 and 0.14 for FIM-PP—improvements of 11 and 21 percentage points, respectively—while approaching the strongest local model (0.38). This represents a significant advance in cross-system generalization. The evolution of tabular foundation models offers a useful parallel: even TabPFNv2 did not consistently outperform strong locally trained methods such as tuned XGBoost on broad benchmarks (e.g., TabArena), yet it established an important research direction. Similarly, FEP demonstrates substantial progress toward general-purpose event prediction, with clear potential for further advances toward—and eventually beyond—locally optimized models.

Finally, we clarify that FEP-Local is exactly the same model as FEP, with identical architecture, hyperparameters, and training procedure; the only difference is that FEP-Local is trained on a single target dataset rather than across heterogeneous datasets. No dataset-specific architecture or hyperparameter optimization is performed for FEP-Local. This controlled comparison is specifically designed to isolate the benefits of cross-dataset pretraining, not to establish practical competitiveness against locally optimized models. The latter is evaluated separately against strong, independently trained and tuned baselines, including Transformer, GRU, Neural Hawkes, and others. Thus, our evidence for FEP's practical competitiveness does not rely on its comparison with FEP-Local.

Weakness 2: Missing nearest-neighbor and fine-tuned foundation-model baselines

R2: We clarify that both nearest-neighbor-based prediction and fine-tuned foundation-model baselines are already included in our evaluation. We also provide additional evidence from a simpler nearest-neighbor predictor.

(1) Nearest-neighbor baselines. FM-v2 (Berti and van der Aalst, 2026b) is itself a nearest-neighbor-based foundation model, which learns representations optimized for retrieving similar prefixes from the target event log. This is explicitly discussed in the Related Work (Lines 119–127), and FM-v2 is included in our experiments, where FEP consistently outperforms it (Table 1). Furthermore, in preliminary experiments, we implemented a simple nearest-neighbor predictor using FEP's learned embeddings instead of its cross-attention-based prediction. This consistently yielded worse results. For example, event-type accuracy dropped from 0.861 to 0.819 on ClinicalTrials.gov, 0.578 to 0.416 on Berka Banking, and 0.606 to 0.563 on EPA ECHO Enforcement. These results further demonstrate the advantage of FEP's learned nonparametric aggregation over simple nearest-neighbor lookup. We will include these additional results in the manuscript.

(2) Fine-tuned foundation-model baseline. Our evaluation already includes FIM-PP (FT) (Berghaus et al., ICLR 2026), which fine-tunes the pretrained FIM-PP on each target dataset using its official adaptation procedure (Table 1, 2 and Appendix A). We report both its zero-shot and fine-tuned performance, explicitly allowing comparison with FEP's no-update approach. Fine-tuning substantially improves FIM-PP, yet FEP still outperforms its fine-tuned variant without any parameter updates. For example, on hard datasets, FIM-PP improves from 0.14 to 0.32 accuracy after fine-tuning, but remains below FEP's zero-shot accuracy of 0.35. Thus, our experiments already examine whether target-specific adaptation of an existing foundation model can close the gap with FEP.

Weakness 3: Missing related work and pretrain-then-fine-tune baselines

R3: We thank the reviewer for highlighting these valuable references, which we will cite and discuss in our paper. However, we emphasize that these works address different learning and transfer settings from FEP, and our experimental evaluation is designed to assess the practical competitiveness of zero-shot inference against strong locally trained models.

(1) Distinction from prior pretraining and LLM-based approaches. Shou et al. (2024) propose self-supervised pretraining with masked and inserted void events, followed by target-specific fine-tuning. Their work focuses on homogeneous transfer, where source and target datasets share the same event types, although their underlying temporal dynamics may differ. CoLES (Babaev et al., 2022) learns self-supervised contrastive representations of event sequences to improve downstream supervised tasks, rather than performing zero-shot next-event prediction. TPP-LLM (Liu and Quan, 2025) leverages textual event descriptions and parameter-efficient LLM fine-tuning for event modeling, while LAMP (Shi et al., 2023) employs LLM-based abductive reasoning to enhance event-sequence prediction. None of these works directly addresses FEP's central challenge: a single pretrained model performing next-event prediction on previously unseen, heterogeneous event systems with unrelated event vocabularies and potentially different temporal dynamics, without any parameter updates. Importantly, FEP requires no additional semantic information, such as textual event descriptions, or prior knowledge of relationships or correspondences between different event systems. We will clarify these distinctions in Related Work.

(2) Scope and rationale of baseline comparisons. Our primary experimental goal is to evaluate how well FEP's zero-shot predictions compare with strong models individually trained and optimized for each target dataset. Accordingly, we include GRU and Transformer, widely used and competitive baselines in predictive process monitoring, together with Neural Hawkes, Transformer Hawkes, and IFTPP, representative established neural temporal point process models. These methods cover different modeling paradigms and undergo dataset-specific training and hyperparameter selection. While additional pretraining and adaptation approaches are valuable, an exhaustive comparison with every locally trained or fine-tuned method is neither feasible nor necessary to evaluate FEP's central contribution.

Moreover, as clarified in R2, our evaluation already includes a pretrain-then-fine-tune baseline, FIM-PP (FT) (Berghaus et al., ICLR 2026), following its official target-specific adaptation procedure. FEP achieves consistently better performance without any target-specific parameter updates. Our goal is not to establish that fine-tuning can never improve performance, but to demonstrate that strong cross-system generalization is achievable without requiring it.

Reviewer_UnNaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an architecture, Foundation Event Predictor (FEP), which addresses the generalization problem across different event logs and datasets, differing in event semantics and patterns. A shared causal transformer encodes context histories as keys and query histories as queries, with the observed next context events as values. It applies cross-attention to predict the next event type and inter-arrival time. Additionally, it uses permutation training or random fixed embeddings for event IDs. The model is trained on 696 real-world logs and evaluated zero-shot on 50 logs. It beats existing foundation-model baselines but has similar or worse performance than local models like Transformer and GRU.

Critical strengths
  • The paper is well written and easy to follow.
  • It addresses an important and practical problem: generalizing across event systems with unrelated event types, patterns.
  • The evaluation is extensive: 50 test datasets and 696 pretraining datasets.
Critical weaknesses
  • Reproducibility: The training code, processed data, and even identities of the 696 pretraining datasets are not shared. Since the datasets come from public sources, the "proprietary constraints" justification is unclear.
  • Missing experimental details: The paper doesn't report pretraining compute. There is no full per-dataset results table.
  • Performance versus local models: FEP doesn't match the local models. This might be expected, but then why would the risk-sensitive applications choose the foundational model over a local one? The paper should identify the datasets where local models clearly outperform FEP and explain why. Per-dataset results are shown for only six datasets.

The paper's main contribution is a pretrained model, but neither the model, the code, nor the list of pretraining datasets is shared. Without them, the community cannot verify, reproduce, or improve the results.

Questions for authors

Please check the weaknesses section.

5 comments

Weakness 1: Reproducibility — Unreleased training code and pretraining datasets

R1: We respectfully disagree with the implication that the absence of released pretraining datasets and training code makes our work scientifically unverifiable or undermines its contributions. We believe this criticism conflates exact reproduction of a particular pretrained model with reproducibility of the underlying methodology. We clarify two important points.

(1) Releasing the complete pretraining data and code is not a prerequisite for disseminating valuable scientific contributions. It is not uncommon for foundation-model research to be published without releasing its complete pretraining corpus or training implementation, even when the underlying data originates from publicly accessible sources. A prominent example is GPT-3 (Brown et al., NeurIPS 2020). Such work can still contribute important architectures, learning paradigms, and methodological insights that advance the research community. It would be unreasonable to dismiss these scientific contributions solely because the original pretraining artifacts cannot be publicly released. Furthermore, the public availability of source datasets does not automatically eliminate legal or proprietary restrictions on disclosing a particular curated training collection or its implementation. We have explicitly acknowledged these restrictions in our Reproducibility Statement.

(2) Our paper provides extensive methodological and experimental details to support independent reproduction and evaluation. Specifically, Section 3 and Figure 1 provide the complete model design and mathematical formulation; Appendix A specifies the architecture, hyperparameters, training configurations, and implementation details, including the complete pretraining procedure in Algorithm 1. Section 4.1 documents the evaluation protocol and baseline selection, while Appendix C provides the pretraining data sources, curation and preprocessing procedures, comprehensive dataset statistics and distributions (Table 9 and Figure 5), and a detailed catalog of every test dataset (Table 10). Together, these details provide a concrete basis for independently implementing FEP, constructing suitable pretraining collections, and evaluating its effectiveness on publicly available benchmarks.

We acknowledge that the exact original pretrained checkpoint cannot be reproduced without the unreleased artifacts. However, this limitation should not be conflated with an inability to reproduce, independently evaluate, or extend the proposed methodology. We believe it is important to distinguish these issues rather than treating the absence of released training artifacts as grounds for dismissing the paper's scientific contributions.

Weakness 2: Missing experimental details — pretraining compute and per-dataset results

R2: We respectfully disagree that the paper lacks sufficient experimental details. The submitted paper already provides extensive documentation of the experimental setup, pretraining procedure, and evaluation results. We address the two specific concerns below.

(1) Pretraining compute and experimental setup. Appendix A explicitly reports the training hardware: four NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, together with the CPU and memory configurations. It also specifies the model architecture, batch sizes, optimizer, learning rate, number of training episodes, distributed training strategy, and other implementation details. Furthermore, Algorithm 1 provides the complete episodic pretraining procedure. If the reviewer is referring specifically to wall-clock training time, pretraining required approximately 125 hours on the reported hardware. We are happy to add this number to the paper.

(2) Per-dataset experimental results. Figure 2 already provides per-dataset comparisons between FEP and the strongest locally trained model across all 50 test datasets, for both event-type accuracy and NMAE. This visualization directly reveals the performance gaps, dataset difficulty, and overall consistency of FEP relative to the best local models. We consider this presentation more informative and space-efficient than a lengthy numerical table for conveying the main findings. Furthermore, Table 8 in Appendix B reports detailed numerical comparisons across all baseline methods on six representative datasets, covering different difficulty levels. Figure 4 additionally provides per-dataset comparisons between FEP and FEP-Local, demonstrating the broad benefits of cross-dataset pretraining.

While we believe the existing presentation already provides clear and comprehensive evidence supporting our conclusions, we are happy to provide the full per-dataset results table for completeness. However, we emphasize that the absence of an exhaustive numerical table should not be interpreted as an absence of per-dataset analysis, which is already explicitly presented in Figures 2 and 4.

Weakness 3: Performance versus local models and practical significance

R3: We respectfully disagree that FEP's slightly lower performance than individually optimized local models undermines its practical value or scientific contribution. We clarify three important points.

(1) The purpose of FEP-Local. FEP-Local is an architecture-matched ablation designed to isolate the benefits of cross-dataset pretraining, rather than an optimized local competitor. It uses the same architecture, hyperparameters, and training procedure as FEP, but is trained separately on each target dataset, without dataset-specific architecture or hyperparameter search. In contrast, Transformer and GRU are individually trained and optimized for each dataset, with careful architecture and hyperparameter selection based on validation performance. Therefore, the performance gap between FEP-Local and these optimized baselines should not be interpreted as evidence of an inherently weaker architecture. The substantial improvement from FEP-Local to pretrained FEP directly demonstrates the benefits of cross-dataset pretraining.

(2) Practical value and research significance. FEP consistently outperforms existing foundation models for event prediction, including FIM-PP and FM-v2, while approaching the performance of carefully optimized local models without any target-specific parameter adaptation. Averaged across the three difficulty groups in Table 1, FEP trails the strongest local baselines by only approximately 2 percentage points in event-type accuracy, while its NMAE is only approximately 0.003 higher. These relatively small average differences must be weighed against the substantial practical benefits of eliminating dataset-specific model training and hyperparameter optimization, reducing ML infrastructure and maintenance costs, alleviating the need for specialized ML expertise, and simplifying deployment under privacy and data-governance constraints. These advantages are particularly valuable when deploying predictive models across numerous heterogeneous event systems. While certain risk-sensitive applications may reasonably prioritize the highest achievable dataset-specific accuracy, this does not negate FEP's advantages in other deployment settings.

More fundamentally, the goal of foundation-model research is not necessarily to outperform every individually optimized model from the outset, but to develop transferable predictive capabilities that eliminate repeated task-specific training and adaptation. This is also reflected in the development of tabular foundation models such as TabPFN, where locally optimized traditional methods can remain competitive in particular settings. Event foundation models remain an emerging research direction. Achieving performance close to strong local models while substantially outperforming existing foundation-model baselines represents a meaningful scientific advance. Please also see our response to Reviewer Yj2w (Weakness 1) for further discussion.

(3) Dataset-wise performance differences. Figure 2 already compares FEP against the best locally trained model on every test dataset, revealing generally small performance gaps alongside a few cases with larger differences. We have not identified a consistent dataset category or characteristic that systematically explains these larger gaps; dataset-specific training and optimization likely contribute to the local models' advantage. As explained in R2, we are happy to provide the complete numerical results table for additional transparency. However, attributing individual performance gaps to specific causes without supporting evidence would be speculative.

Overall, FEP demonstrates that a single pretrained model can generalize across heterogeneous event systems and achieve performance close to carefully optimized local models without target-specific training. We believe this is a significant advance in both methodology and practical applicability, rather than a weakness of the proposed approach.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Reminder to Respond to Rebuttals

We would like to thank the reviewers for their time assessing our paper. We politely ask the reviewers to respond to our rebuttals so we can conclusively address their concerns before the discussion period ends tomorrow, November 18.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.